JAMRIS 2009 Vol 3 No 2 by JAMRIS

JOURNAL of AUTOMATION, MOBILE ROBOTICS & INTELLIGENT SYSTEMS

Editor-in-Chief Janusz Kacprzyk

Executive Editor: Anna Ładan aladan@piap.pl

(Systems Research Institute, Polish Academy of Sciences; PIAP, Poland)

Associate Editors: Mariusz Andrzejczak (PIAP, Poland) Katarzyna Rzeplińska-Rykała (PIAP, Poland)

Co-Editors: Dimitar Filev (Research & Advanced Engineering, Ford Motor Company, USA)

Kaoru Hirota

Webmaster: Tomasz Kobyliński tkobylinski@piap.pl

(Interdisciplinary Graduate School of Science and Engineering, Tokyo Institute of Technology, Japan)

Witold Pedrycz

Proofreading: Urszula Wiączek

(ECERF, University of Alberta, Canada)

Roman Szewczyk (PIAP, Warsaw University of Technology, Poland)

Editorial Office: Industrial Research Institute for Automation and Measurements PIAP Al. Jerozolimskie 202, 02-486 Warsaw, POLAND Tel. +48-22-8740109, office@jamris.org

Editorial Board: Chairman: Janusz Kacprzyk (Polish Academy of Sciences; PIAP, Poland) Plamen Angelov (Lancaster University, UK) Zenn Bien (Korea Advanced Institute of Science and Technology, Korea) Adam Borkowski (Polish Academy of Sciences, Poland) Wolfgang Borutzky (Fachhochschule Bonn-Rhein-Sieg, Germany) Oscar Castillo (Tijuana Institute of Technology, Mexico) Chin Chen Chang (Feng Chia University, Taiwan) Jorge Manuel Miranda Dias (University of Coimbra, Portugal) Bogdan Gabryś (Bournemouth University, UK) Jan Jabłkowski (PIAP, Poland) Stanisław Kaczanowski (PIAP, Poland) Tadeusz Kaczorek (Warsaw University of Technology, Poland) Marian P. Kaźmierkowski (Warsaw University of Technology, Poland) Józef Korbicz (University of Zielona Góra, Poland) Krzysztof Kozłowski (Poznań University of Technology, Poland) Eckart Kramer (Fachhochschule Eberswalde, Germany) Andrew Kusiak (University of Iowa, USA) Mark Last (Ben–Gurion University of the Negev, Israel) Anthony Maciejewski (Colorado State University, USA) Krzysztof Malinowski (Warsaw University of Technology, Poland)

Andrzej Masłowski (PIAP, Poland) Tadeusz Missala (PIAP, Poland) Fazel Naghdy (University of Wollongong, Australia) Zbigniew Nahorski (Polish Academy of Science, Poland) Antoni Niederliński (Silesian University of Technology, Poland) Witold Pedrycz (University of Alberta, Canada) Duc Truong Pham (Cardiff University, UK) Lech Polkowski (Polish-Japanese Institute of Information Technology, Poland) Alain Pruski (University of Metz, France) Leszek Rutkowski (Częstochowa University of Technology, Poland) Klaus Schilling (Julius-Maximilians-University Würzburg, Germany) Ryszard Tadeusiewicz (AGH University of Science and Technology in Kraków, Poland)

Stanisław Tarasiewicz (University of Laval, Canada) Piotr Tatjewski (Warsaw University of Technology, Poland) Władysław Torbicz (Polish Academy of Sciences, Poland) Leszek Trybus (Rzeszów University of Technology, Poland) René Wamkeue (University of Québec, Canada) Janusz Zalewski (Florida Gulf Coast University, USA) Marek Zaremba (University of Québec, Canada) Teresa Zielińska (Warsaw University of Technology, Poland)

Publisher: Industrial Research Institute for Automation and Measurements PIAP

If in doubt about the proper edition of contributions, please contact the Executive Editor. Articles are reviewed, excluding advertisements and descriptions of products. The Editor does not take the responsibility for contents of advertisements, inserts etc. The Editor reserves the right to make relevant revisions, abbreviations and adjustments to the articles.

JOURNAL of AUTOMATION, MOBILE ROBOTICS & INTELLIGENT SYSTEMS VOLUME 3, N° 2, 2009

CONTENTS SPECIAL ISSUE SECTION

REGULAR PAPERS 59

AH/EHW - the State of the Art and the Prospectus for Future Developments Editors: Mircea Gh. Negoita, Sorin Hintea

Dynamical approach to the diagonal gait synthesis: theory and experiments T. Zielińska, M. Trojnacki

8 Computation of positive realizations of siso singular hybrid linear systems T. Kaczorek, Ł. Sajewski

EDITORIAL M.Gh. Negoita, S. Hintea

15 An effective localization method for robot navigation through combined encoders positioning and retiming visual control M. Jallouli, L. Amouri, N. Derbel

Analog circuit design based on computational intelligence techniques G. Oltean, S. Hintea, E. Sipos

70 Adaptive and evolvable hardware and systems: the state of the art and the prospectus for future development M.Gh. Negoita, L. Sekanina, A. Stoica

24 On path planning for mobile robots: introducing the mereological potential field method in the framework of mereological spatial reasoning P. Ośmiałowski

76 A highly linear low pass filter for low voltage reconfigurable wireless applications G. Csipkes, S. Hintea, D. Csipkes, C. Rus, L. Festila, H. Fernández-Canque

34 An encoded infrared sheet of light navigational beacon system for precise localization of indoor mobile robot vehicles J. Kurian, P.R. Saseendran Pillai

82 An analog linear SVM image classifier L. Festila, L.A. Szolga, R. Groza, S. Hintea, M. Cirlugea

42 Proper selection of thermal insulation materials T.M. El-Shiekh, A.A. Elsayed

88 I/Q Imbalance compensation algorithm based on neural networks B.S. Kirei, M. Topa, M. Neag, R.C. Onet

47 A new design of a robot prototype for intelligent navigation and parallel parking C. Abdelmoula, F. Chaari, M. Masmoudi

DEPARTMENTS 93

IN THE SPOTLIGHT 94

EVENTS 2

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

NÂ° 2

2009

DYNAMICAL APPROACH TO THE DIAGONAL GAIT SYNTHESIS: THEORY AND EXPERIMENTS Received 6th May 2008; accepted 29th September 2008.

Teresa ZieliĹ&#x201E;ska, Maciej Trojnacki

Abstract: The method of two-legged diagonal gait synthesis for the quadruped robot is introduced. The problem of dynamic postural equilibration taking into account the role of compliant feet is solved. The equilibrium conditions are split to the feet attachments points and the points within the feetend area. In simulation example the slip avoidance condition is tested. Presented method has the meaning for motion synthesis taking into account the robot parameters and - for the design of the feet considering the dynamically stable gaits. The method was proved by using simulations and experiments.

fingers and two active DOFs - Fig. 3 is the unique [6] example of the more complex structure. In gait synthesis the attention is paid to the positioning of active joints. The role of foot and its passive DOFs during the walk of multi-legged machine is often neglected. On the other hand the usefulness of passive joints and springs in walking machines have been confirmed by practical experience. In this article we discuss the role of passive joints in maintaining the postural dynamical stability. Theoretical considerations are supported by simulation, and the results were validated by experiments.

2. Problem statement Keywords: quadruped robot, diagonal gait, postural equilibrium, force distribution, computer simulation.

1. Introduction Many works concerning multi-legged walking machines are devoted to the design problems and to the motion generation principles of statically stable locomotion (i.e. [1, 3, 4, 5, 7, 10]). The control aspects are also discussed [14, 15, 16]. Our thorough search bringed brought no publications on the analysis of equilibrium conditions for the dynamical gaits considering stabilizing role of the foot.

The foot considered in our work is illustrated by the Fig. 4. In prismatic joint connecting the foot with the shank the spring is mounted. This brings the leg compliance, because the spring length changes proportionally to the vertical force. The length change is small but it supports the postural equilibrium, as it will be discussed further.

Fig. 2. Leg-end joint with 3 DOFs.

Fig. 1. Foot shaped as a ball or as a plate. The legs of multi-legged walking machines have typically 2 or 3 active degrees of freedom. The additional degrees of freedom (if introduced) are passive. The foot compliance is typically obtained by using the springs. Many multi-legged robots have feet shaped as balls or as a rotating plates [13].They are attached to the shank by passive prismatic joints - Fig. 1. More complex designs consist of 3 passive DOFs [2]. The potentiometers are sometimes utilized as sensors for monitoring the joints position - Fig. 2. The biologically inspired foot with three

Fig. 3. Scheme of the foot inspired by the animals build. Let's assume that OXYZ is non-moving reference frame, RXYZ is the frame attached to the robot trunk. Coordinates of the robot mass centre are Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

equal to: (1) where

mass of trunk, total mass. To illustrate the method we will consider the prototype quadruped with dimensions and masses distribution illustrated in Fig. 4 (dimensions are expressed in mm). Lower script si denote the supporting leg. In considered diagonal gait the supporting legs are 1 and 4, or 2 and 3, where 1 denotes left front leg, 2 - is the right front leg, 3 - left hind leg, 4 right hind leg. In further considerations s1 and s2 will mark the pair of supporting legs (for s1 =1 is s2 =4, and for s1 =2 is s2 =3). Force equilibrium conditions are usually expressed in external reference frame OXYZ. They can be easily transformed to the local frame RXYZ [12]. In frame RXYZ the conditions are expressed as:

N째 2

2009

is transformation matrix from OXYZ to RXYZ. is the resultant force vector acting to the robot mass centre. is the force vector exerted by leg-end. The torques equilibrium conditions are expressed in frame RXYZ by:

(3) are the external moments applied to the robot. In our considerations we assume only the rotation along vertical axis Z passing point R, is the main inertia moment around axis Z, and is the rotation angle. With the above assumption it isNNNNNNNNNNNNNNN . Shortening, we denote by .

(2) (4) where

is the gravity constant.

Fig. 4. View of the robot structure, its foot, and illustration of forces.

Fig. 5. Robot foot and applied notation, view of the prototype. 4

Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

N째 2

2009

where (8)

(5)

where is the point height over the ground considering spring shortening under the vertical force . We keep in mind that . Now we define - intersection point of with supporting plane (Fig. 4). This point is also the attachment point of resultant reaction force vector. During the real (physical) walk it is also the Centre Of Pressure (COP). In smooth walk (with smooth motion of the trunk) it is expected that . Point COP being the intersection of vector with plane has coordinates:

(6) (9) Matrix is singular (rank( ), rank of extended matrix is 6 what means that equalities can not be fulfilled. The equilibrium conditions described by (2), (3) cannot be fulfilled considering only the legends. Taking into account the stabilizing role of the feet we split the conditions between points and of supporting legs. Passive joint in foot attachment allows the rotation along axis parallel to Y, but not rotation along axis parallel to X. Therefore, we can consider that to prevent the side trunk inclination the moment is compensated by forces exerted in points of supporting legs. For evaluation of vertical leg-end forces we assume that the points are located in constant distance from the plane RXY. Taking into account , the forces equilibrium conditions, and condition for equilibrating nd th the moment (2 and 4 equality from rel. (4)) we obtain:

(10) In stable posture the moments results from reaction forces, and evaluated in supporting plane towards the reference point COP shall be equal to zero:

(11)

(12) rd

Now, considering (11) and (12), and 3 equality from (4) we obtain:

(7) Now the forces are known. Using (3) the moments for both supporting feet can be easily obtained. Those moments must be compensated by feet compliance. We keep in mind that the rotations are only possible in passive joints located in points (Fig. 5). The springs mounted in feet are bending under the exerted vertical leg-end forces. The active range of springs bend must be chosen to match the range of vertical forces. The springs bend introduces the stabilizing effect. Forces acting on mounting points are accordingly transferred to the points which are application points of reaction force vectors . Points are translated in plane (Fig. 5), that means XY in relation to etc. The shifts dxsi, dysi are such that the moments resulting from reaction forces towards compensate the moments .

(13)

where . Remembering that using the expressions (7) the forces are also obtained, we rearrange each equation (13) separately, to obtain:

Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

N째 2

2009

due to the shift of leg-end force vector within the sole area, as has been discussed in our work.

(14) Substituting (14) to (8) and rearranging the terms we have:

(15) The moments equilibrated by feet can be obtained expressing (2), (3) for each foot separately, that means:

(16) After calculation of those moments, the relations (15) will be applied to calculate first and then, on a basis of (14), values of will be obtained. Now the relations (8) can be used to bring the remaining force components (besides of already calculated vertical forces ).

3. Simulation research Robot structure and its mass distribution are given in Fig. 4. Leg has 3 active DOFs -2 in the hip and 1 in the knee. As it was mentioned shank and foot are connected by passive joint. The attachment uses in-build vertical spring (Fig. 5). Robot and its control system was described in [8, 9], in [11] the motion properties were discussed. The robot trunk length is , width , thigh and shank lengths . Total mass , trunk mass , mass of hip segment , mass of thigh , mass of shank , foot masses: the upper part , the lower part . Those parameters were used in simulations. The results obtained for robot turning motion with rotation by are presented. The robot height was . Results are given in Fig. 6. The minimal friction coefficient expressed as was also evaluated. Feet will not slip when the real coefficient is not smaller than . The experimental confirmation of the feet role in postural stabilization is given by observed stable displacement of the device by diagonal gait. Having the dismounted feet machine will overturn when trying to move by that gait. This proves that the stable posture is obtained 6

Articles

Fig. 6. Coordinates of COP, reaction forces and minimum friction coefficient.

Fig. 7. Registered current in knee motor i3i (motor no.3, i is the leg number).

Fig. 8. Calculated torques for knee joint number).

( is the leg

4. Conclusion The motors current was monitored during the real gait. The current is proportional to the actuating torque. The torques was also evaluated using the inverse dynamics model and considering the forces obtained by the method discussed above. The measured current

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

(Fig. 7) and calculated torques (Fig. 8) exhibited similar regularities that prove the correctness of the presented method for leg-end forces evaluation. The confirmation of the feet importance for postural stabilization was the stable walk observed for diagonal gait. The presented method of forces evaluation decomposes equilibrium conditions considering the feet attachments and feet-ends. Neglecting the condition for the rotation around daxis Z was allowed. In considered prototype (as in majority of quadrupeds) the feet are rotating freely, therefore the assumption has a logic confirmation. Postural stability observed during real machine motion by diagonal gait confirms the correctness of presented considerations. Obtained relations are useful for the walking machines design. The knowledge of leg-end forces translation towards the feet mounting points has the importance for synthesis of dynamical stable gaits. The knowledge of those translations is necessary to evaluate if the foot supporting area will assure the postural stability. Relying on the given results the change the foot area or change of the leg configuration (changing the position of CG) can be adjusted for postural stabilization.

[6]

[7]

[8]

[9]

[10]

[11]

ACKNOWLEDGMENTS The work on this article was supported by University Research Program (UPB) and Ministry of Scientific Research and Information Technology Research Grant - Analysis of Stabilization Mechanisms for Two-legged Locomotion.

[12]

[13]

AUTHORS Teresa Zielińska* - Profesor at the Institute of Aeronautics and Applied Mechanics (WUTIAAM), Warsaw University of Technology, ul. Nowowiejska 24, 00-665 Warsaw, Poland. E-mail: teresaz@meil.pw.edu.pl. Maciej Trojnacki - Industrial Research Institute for Automation and Measurements (PIAP), Al. Jerozolimskie 202, 02-486 Warsaw, Poland. E-mail: mtrojnacki@piap.pl. * Corresponding author

[14]

[15]

[16]

References Gardner J.F., “Efficient Computation of Force Dis[1] tribution for Walking Machines on Rough Terrain”, Robotica, vol. 10, no. 5, 1992, pp. 427-433. [2] Garcia E., Galvez J.A., Gonzalez de Santos P., “On Finding the Relevant Dynamics for Model-Based Controlling Walking Robots”, Journal of Intelligent and Robotic Systems, vol.37, issue 4, 2003, pp. 375 -398. [3] Klein C.A, Kittivatcharapong S., “Optimal Force Distribution for the Legs of a Walking machine with Friction Cone Constraints”, IEEE Trans. on Robotics and Automation, vol. 6, no. 1, 1990, pp. 73-85. [4] Martins-Filho L.S., Prajoux R., “Locomotion Control of a Four-legged Robot Embedding Real-time Reasoning in the Force Distribution”, Robotics and Autonomous Systems, vol. 32, 2000, pp. 219-235. [5] Pfeiffer F., Eltze J., Weidemann H.J., “Six-legged

N° 2

2009

Walking Considering Biological Principles”, Robotics and Autonomous Systems, vol. 14, 1995, pp. 223-232. Spenneberg D., Albrecht M., Backhaus T., Hilljegerdes J., Kirchner F., Strack A., and Schenker H., “Aramies: A four-legged Climbing and Walking Robot”. In: Proc. of 8th International Symposium iSAIRAS, Munich, September 2005 (CD ROM). Takemura H., Deguchi M., Ueda J., Matsumoto Y., Ogasawara T., “Slip-adaptive Walk of Quadruped Robot”, Robotics and Autonomous Systems, vol. 53, 2005, pp. 124-141. Trojnacki M., „Motion Description of Quadruped Robot”, Zeszyty Naukowe Politechniki Rzeszowskiej, no. 222, Rzeszów 2005, pp. 357-364 (in Polish). Trojnacki M., “The Modeling, Programming and Computer Simulation of Motion for a Four-legged Robot”. In: Projektowanie Mechatroniczne, Ed. T. Uhl, Wydawnictwo Instytutu Technologii Eksploatacji PIB: Kraków 2006, pp. 149-160. Zhou D., Low K.H., Zieliska T., “An Efficient Footforce Distribution Algorithm for Quadruped Walking Robots”, Robotica, vol.18, 2000, pp. 403-413. Zielińska T., Trojnacki M., “Motion Synthesis of Dynamically Stable Two-legged Gait for a Quadruped Robot. Theoretical Considerations (1)”, PAR, no. 11, 2007, pp. 5-11 (in Polish). Zielińska T., Walking Machines: Fundamentals, Design, Control and Biological Patterns, PWN: Warsaw, 2003, (in Polish). Zielińska T., Heng J., “Mechanical Design of Multifunctional Quadruped”, Mechanism and Machine Theory, vol.38, 2003, pp. 463-478. Zielińska T., Motion Synthesis. In Walking: Biological and Technological Aspects, CISM Courses and Lectures no. 467. Ed. by Pfeiffer F., Zielinska T., Springer Verlag, 2004, pp. 151-187. Zielińska T., “Control and Navigation Aspects of a Group of Walking Robots”, Robotica, Cambridge University Press, vol.24, 2006, pp. 23-29. Zielińska T., Heng J., “Real-time Control System for a Group of Autonomous Walking Robots”, Advanced Robotics, vol.20, no.5, 2006, pp. 543-561.

Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

Nﾂｰ 2

2009

COMPUTATION OF POSITIVE REALIZATIONS OF SISO SINGULAR HYBRID LINEAR SYSTEMS Received 28th April 2008; accepted 10th September 2008.

Tadeusz Kaczorek, ﾅ「kasz Sajewski

Abstract: The realization problem for 2D positive singular linear hybrid systems is formulated, as well as a method based on the state variable diagram for finding a positive realization of a given improper transfer function is proposed. Sufficient conditions for the existence of a positive realization of a given improper transfer function are established. A procedure for computation of a positive realization is proposed and illustrated by a numerical example.

2. Preliminaries and formulation of the problem Consider a SISO singular hybrid system described by the equations

(1a)

(1b)

Keywords: hybrid linear systems, singular, positive, realization. where

1. Introduction In positive systems inputs, state variables and outputs take only non-negative values. Examples of positive systems are industrial processes involving chemical reactors, heat exchangers and distillation columns, storage systems, compartmental systems, water and atmospheric pollution models. A variety of models having positive linear systems behaviour can be found in engineering, management science, economics, social sciences, biology and medicine, etc. Positive linear systems are defined on cones and not on linear spaces. Therefore, the theory of positive systems is more complicated and less advanced. An overview of the state of the art in positive systems theory has been given in the monographs [2, 9]. The realization problem for positive discrete-time and continues-time systems without and with delays was considered in [1, 2, 9-13]. The reachability, controllability and minimum energy control of positive linear discrete-time systems with delays have been considered in [3]. The relative controllability of stationary hybrid systems has been investigated in [20] and the observability of linear differentialalgebraic systems with delays has been considered in [21]. A new class of positive 2D hybrid linear system has been introduced in [14]. The realization problem for this class of systems has been considered in [6, 15] and for class of positive hybrid systems with delays in [5, 7]. The main purpose of this paper is to present a new method for computation of positive realizations for 2D single-input single-output singular hybrid linear systems using the state variable diagram method. Sufficient conditions for the existence of a positive realization of a given improper transfer function will be established and a procedure for computation of positive realizations will be proposed. Considerations will be illustrated by numerical example.

Articles

are the state vectors and output vectors and

are input are

real matrices. It is assumed that

and

Boundary conditions for (1a) have the form and

(2)

Note that the hybrid system (1) has the similar structure as Roesser model [5, 14, 16]. Let is the set of real matrices with nonnegative entries and . Definition 1. The hybrid system (1) is called internally positive if and for arbitrary boundary conditions and all inputs . Transfer matrix of the system (1) is given by the formula (3) Definition 2. The matrices are called the positive realization of the improper transfer matrix, if they satisfy the equality (3). A realization is called minimal if the matrices and have minimal dimensions among all positive realizations of . The realization problem can be stated as follow. With a given proper rational transfer function , find its positive realization .

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

N째 2

2009

3. Problem solution The essence of proposed method for solving of the realization problem for positive 2D hybrid systems will be presented on the following improper transfer function (4) Multiplying the numerator and denominator of transfer function (4) by

we obtain (5)

Defining (6) from (5) and (6) we obtain (7)

Using (7) we may draw the state variable diagram shown in Fig.1

Fig. 1. State variable diagram for transfer function (4). As a state variable we choose outputs of integrators and of delay elements . We are dealing with the singular system and we have to choose one more state variable in this case . Using the state variable diagram (Fig.1) we can write the following state equations

(8) Taking into account Definition 1 and (8), the following has been proved. Theorem 1. The singular hybrid system with improper transfer function (4) is positive if the following conditions are Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 2

2009

satisfied: all coefficients of numerator of improper transfer function are non-negative, all coefficients of denominator of improper transfer function are non-negative. Rewriting equations (8) in matrix form

(9) and if conditions of Theorem 1 are met, then positive realization of improper transfer function (4) of singular hybrid system (1) has the form:

(10) Generalizing the considerations on any improper transfer function we obtain

(11) where

Multiplying the numerator and denominator of transfer function

we obtain (12)

Defining

(13) we may draw the state variable diagram shown in Fig.2 10

Articles

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 3,

N째 2

2009

Fig. 2. State variable diagram for transfer function (11). As a state variable we choose the outputs of integrators and of delay elements . Using state variable diagram (Fig. 2) we can write the following differential and difference equations

(14) Defining vectors

(15) equations (14) we may write in the form

(16) Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

N째 2

2009

where

(17) Therefore, the following theorem has been proved. Theorem 2. There exists a positive realization of the improper transfer function (11) if all coefficients of its numerator and denominator are nonnegative. are negative. Remark. A characteristic feature of singular positive 2D hybrid model is that some entries of the matrix If the assumptions of Theorem 2 are satisfied then a positive realization can be found by the use of the following procedure. Procedure. in the form (12) and the equations (13), Step 1. Write the transfer function Step 2. Using (13) draw the state variable diagram shown in Fig. 2, Step 3. Choose the state variables and write equations (14), Step 4. Using (14) find the desired realization (17) of the transfer function (11).

4. The example There is given improper transfer function (18) find its positive realization (17). In this case

and

Using the procedure we obtain the following: Step 1. Multiplying the numerator and denominator of transfer function (18) by 12

Articles

we obtain

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 2

2009

5. Concluding remarks (19) and

(20) Step 2. In this case state variable diagram has the form shown in Fig. 3

A method for computation of a positive realization of a given improper transfer function of 2D hybrid linear system has been proposed. Sufficient conditions for the existence of a positive realization of a given improper transfer function have been established. A procedure for computation of a positive realization has been proposed, and illustrated by numerical example. An open problem is formulation of the necessary and sufficient conditions for the existence of solution of the positive realization problem for 2D hybrid systems in the general case. Extensions of those considerations for 2D hybrid systems described by models with structures similar to the 2D general model or the 2D second Fornasini-Marchesini model [9, 14] are also open problems. ACKNOWLEDGMENTS This work was supported by the State Committee for Scientific Research of Poland under the grant NN 514 1939 33.

AUTHORS Tadeusz Kaczorek and Łukasz Sajewski* - Bialystok Technical University, Faculty of Electrical Engineering, Wiejska 45D, 15-351 Bialystok. E-mail: lsajewski@doktoranci.pb.edu.pl * Corresponding author

References Fig. 3. State variable diagram for transfer function (18). Step 3. Using the state variable diagram (Fig. 3) we can write the state equations

[1]

[2] [3]

[4]

(21)

Benvenuti L., Farina L., “A tutorial on the positive realization problem”, IEEE Trans. Autom. Control, vol. 49, no. 5, 2004, pp. 651-664. Farina L., Rinaldi S., Positive Linear Systems; Theory and Applications, J. Wiley: New York, 2000. Kaczorek T., Busłowicz M., “Reachability and minimum energy control of positive linear discrete-time systems with one delay”. In: 12th Mediterranean Conference on Control and Automation, 6th-9th June, 2004, Kusadasi, Izmir, Turkey. Kaczorek T., Busłowicz M., “Minimal realization problem for positive multivariable linear systems with delay”, Int. J. Appl. Math. Comput. Sci., vol. 14, no. 2, 2004, pp. 181-187.

Step 4. Desired positive realization has the form

(22)

Articles

Journal of Automation, Mobile Robotics & Intelligent Systems

[5]

[6]

[7]

[8]

[9] [10]

[11]

[12]

[13]

[14]

[15] [16] [17]

[18] [19]

[20]

[21]

[22]

Kaczorek T., Sajewski Ł., “Computation of positive realization of MIMO hybrid linear systems with delays using the state variable diagram method”. In: 16th International Conference on Systems Science, 4th-6th September, 2007, Wrocław, vol. 1, 2007, pp. 150-160. Kaczorek T., Sajewski Ł., “Computation of positive realization of MIMO hybrid linear systems using the state variable diagram method”, Archives of Control Sciences, vol. 17, no. 1, 2007, pp. 5-21. Kaczorek T., Sajewski Ł., “Realization problem for positive 2D hybrid systems with one delay in state and input vectors”. In: 8th International Workshop „Computational Problems of Electrical Engineering”, 14th-16th September 2007, Wilkasy, Poland, Przegląd Elektrotechniczny, no. 2, 2007, pp. 242-246. Kaczorek T., “Some recent developments in positive systems”. In: Proc. of 7th Conference of Dynamical Systems Theory and Applications, Łódź 2003 pp. 25-35. Kaczorek T., Positive 1D and 2D systems, Springer Verlag: London 2002. Kaczorek T., “A realization problem for positive continues-time linear systems with reduced numbers of delay”, Int. J. Appl. Math. Comp. Sci., vol. 16, no. 3, 2006, pp. 325-331. Kaczorek T., “Realization problem for positive multivariable discrete-time linear systems with delays in the state vector and inputs”, Int. J. Appl. Math. Comp. Sci., vol. 16, no. 2, 2006, pp. 101-106. Kaczorek T., “Realization problem for positive discretetime systems with delay”, System Science, vol. 30, no. 4, 2004, pp. 117-130. Kaczorek T., “Positive minimal realizations for singular discrete-time systems with delays in state and delays in control”, Bull. Pol. Acad. Sci. Techn., vol 53, no 3, 2005, pp. 293-298. Kaczorek T., “Positive 2D hybrid linear systems”. In: Proc. Inter. Conf. Numerical Linear Algebra in Signals Systems and Control 2007. Kaczorek T., “Realization problem for positive 2D hybrid systems”, Submitted to COMPEL. Kaczorek T., Two-Dimensional Linear Systems, Springer Verlag: Berlin 1985. Kaczorek T., Determination of singular positive realization of improper transfer function of 2D linear systems, SMC Zakopane 2007. Klamka J., Controllability of Dynamical Systems, Kluwer Academic Publ.: Dordrecht, 1991. Kurek J., “The general state-space model for a two-dimensional linear digital system”, IEEE Trans. Austom. Contr. AC-30 , June 1985, pp. 600-602. Marchenko V. M., Poddubnaya O. N., “Relative controllability of stationary hybrid systems”. In: 10th IEEE Int. Conf. on Methods and Models in Automation and Robotics, 30th August - 2nd Sept. 2004, Międzyzdroje, Poland, pp. 267-272. Marchenko V. M., Poddubnaya O. N., Zaczkiewicz Z., “On the observability of linear differential-algebraic systems with delays”, IEEE Trans. Autom. Contr., vol. 51, no. 8, 2006, pp. 1387-1392. Roesser R. B., “A discrete state-space model for linear image processing”, IEEE Trans. on Automatic Control AC20, no. 1, 1975, pp. 1-10.

Articles

VOLUME 3,

N° 2

2009

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

N째 2

2009

AN EFFECTIVE LOCALIZATION METHOD FOR ROBOT NAVIGATION THROUGH COMBINED ENCODERS POSITIONING AND RETIMING VISUAL CONTROL Received 7th October 2008; accepted 23rd December 2008.

Mohamed Jallouli, Lobna Amouri, Nabil Derbel

Abstract: The paper presents an optimal mobile robot localisation method through encoders' measurements and absolute localization using webcam data. This technique has been developed and implemented for the motion of the robot from an initial position towards another desired position, taking into account Kinematic constraints. The proposed method is a cheap technique, which deals with the problem of robot initial position as the visual system provides the origin absolute coordinates related to the reference system. First, we have developed an interface, which ensures the real-time localisation of the mini robot Khepera II while tracking a virtual moving target. Secondly, we have carried out experimentations of both the relative localization method, which determines the speed values of each driving wheel, and absolute localization using webcam data, which determines the robot start position. Next, a correction of the mobile robot position has been realized by retiming points using webcam measures. The obtained results are compared and discussed through different trajectories. Keywords: localization, encoders' measurements, fuzzy controller, webcam data, segmentation, retiming method.

1. Introduction During the last decades, the useful range for robots has gradually spread to a wide variety of areas. Mobile robots are especially being used as a substitute for humans or to do simple work that is either in or outside. In such a mobile robot system, getting exact information on its current position is very important. At all times, the mobile robot must know instantaneously its current position and the one of the objective. There are two positionestimation methods applied in navigation systems, i.e. absolute and relative positioning. Relative localization is realised through measurements provided by sensors measuring internal variables of the vehicle. The incremental encoders are the typical inertial sensors. These sensors placed on the wheels axis, which represents the rotation axis of the vehicle. The diesadvantage of this method is that errors of each measurement are accumulated. This heavily degrades the estimates of the position and the orientation of the vehicle, especially for long and winding trajectories [1]. However, absolute localization is based on the use of sensors measuring some parameters of the environment in which the robot is operating. A set of sonars is generally used as an external sensory device. The infrared sensors are implemented on the robot and measure the distance

with the environment. These sensors are also widely utilized for the guidance of autonomous vehicles with obstacle avoidance in unknown environment [2, 3]. The major disadvantage of absolute measurements is their dependence on the characteristics of the environment. In order to compensate these drawbacks, the localization method that fuse data coming from odometers and sonar sensors by applying Hybrid Kalman Filter Fuzzy Logic Adaptive Multisensor Data Fusion Architectures [4, 5] can be used. An other approach adapts the position and the orientation of a mobile robot through a weighted Extended Kalman Filter (EKF) [6, 7]. These methods need much calculation for a mobile robot to perform a task. Other disadvantages are either the short range of used sensors or the necessity to know the initial position of the robot. Other solution uses a method that estimates a position of a robot using a CCD camera fixed on ceiling of the corridor by calculating the distance moved and time of a mobile robot [8]. Another way that is presented for a robot equipped with a CCD camera calculates its position by recognizing a characteristic topography and compares it with the model image saved in memory [9]. A last proposed method calculates the position of the robot in order to intercept a moving target through visual feedback [10]. The most important disadvantage of these methods is the necessity to know the initial position of the robot. In this paper, the proposed method uses two different sensors. First, with encoders, which ensure relative positioning using the kinematic model, a fuzzy controller is used to attempt a virtual moving target. Second, a camera installed on the ceiling of the test environment ensures the absolute localization using a segmentation method and a location algorithm and reduces the encoder position errors regardless of a retiming visual control. The experimental results show the effectiveness of the proposed algorithm. The system compensates for robot positioning by means of the following sequences. First, the system calculates the initial position and the orientation by using webcam data. Secondly, the fuzzy controller generates velocities to be applied on the mobile base during an optimal trajectory from the initial position towards another desired position, taking into account kinematic constraints. During the robot navigation, the system calculates its position using encoders' measurements and webcam data. The system compensates encoder errors by retiming point using the visual controller. This article is organized in the following way. In section 2, we present the relative localization system studied with mathematical formulation of the problem Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

and the optimised fuzzy controller design. Section 3 describes the absolute localization system used with the position and the orientation estimation algorithm. Section 4 shows the experimental tests applied to the mobile base Khepera II.

2. Relative localization In this work, relative localization is realised through measurements provided by the encoders fixed on the mobile robot wheels and used both the kinematic model and the fuzzy controller system. 2.1. The mathematical formulation Several types of mobile robots with driving wheels and encoder system have been studied in the literature [11, 12]. The most studied types are those of the steering angle commanded vehicles [13, 14, 15]. However, our experiments will be carried up onto a mobile robot with two independent driving wheels, which can be oriented and commanded by acting on the speed of each wheel, as shown on the schematic model (Fig. 1).

N째 2

2009

To further optimise the mobile trajectory, we propose a fuzzy controller calculating the velocities to be applied on the robot wheels. 2.2. Fuzzy controller synthesis The most employed strategies for the mobile robot navigation uses Fuzzy Logic Controllers (FLC). The contribution of our study is to optimise a fuzzy controller by means of the gradient method and its validation by the implementation onto the mobile robot Khepera II. The fuzzy controller allows the robot to reach a virtual moving target point starting from a given position. The robot position and orientation are calculated in real time by the odometer module. The controller outputs are velocities and applied respectively on the left and the right driving wheels of the mobile robot, in order to reach a desired position. A bloc diagram of the fuzzy logic controller is shown in Fig. 2.

Fig. 2. Block diagram of the fuzzy logic controller optimised by Gradient Method. The controller has two inputs the distance and the angle . Where is the distance between the centre of the robot and its target; and is the difference between the angle of robot's direction and the angle of the connecting line between the centre of the robot and its target (see Figure 3). The distance d and the angle are expressed as follows: Fig. 1. The schematic model of a nonholonomic mobile robot. (3) The kinematic model is given by: (4) with (1)

(5)

where and are the robot's right and left wheel's velocities, respectively; is the robot's angular velocity, L is the distance between two wheels and is the angle between the robot's direction and the X-axis. By discretization of the system (1) using Euler method, it becomes:

(2)

where T is the sampling time. 16

Articles

Fig. 3. Robot configuration according to the objective.

Journal of Automation, Mobile Robotics & Intelligent Systems

From our experiments, we construct five membership functions for the distance and four for the angle . and are fuzzy subsets defined by their corresponding membership functions, i.e., and . Five fuzzy subsets are assigned to the variable distance : VS: Very Small; S: Small; M: Medium; L: Large; VL: Very Large. Four fuzzy subsets have been associated to the angle : Z: Zero; PS: Positive Small; PM: Positive Medium; PL: positive Large. For each input variable value combination, an action of the output variables is associated to it. Fuzzy rules (situation/action) are proposed in table 1. These rules are manually constructed following several simulations. With the notations: : Left Velocity, : Right Velocity, B: Big, VB: Very Big, S: Small, Z: Zero, M: medium. Table 1. Linguistic inference table.

3. Absolute localization In this proposed method, the absolute localization is based on the use of a webcam, which measures some parameters of the environment in which the robot is operating. This approach is based on four steps. The first one consists on the image pre-processing. The second one is an orientation estimation with the external sensor. In the third step, we propose an optimising pre-processing method in order to reduce the time processing and to ensure

VOLUME 3,

N째 2

2009

real time localization. In the fourth and last step, we present the transformation from image coordinates to real coordinates algorithm. 3.1. Image pre-processing We extract the pixel coordinates of the reference system using region properties of a reference image, which is captured and stored in advance. This image describes the workspace of the mobile robot limited by four landmarks placed on the corners. Figures 4 and 5 describe this reference image. In this paper, we apply a binary mask for removing illumination image noises, and selected images, which have 640*480 pixels for an image. The filtering method that has been adopted is a morphological one. Through labelling, we separate objects and search for their features [16]. Then, to recognize whether object is the robot or not, we use region properties (areas and centroid) of an input image which is being input consecutively. 3.2. Orientation Estimation In this section, we will introduce a technique used in order to calculate the orientation of the mobile robot. We have attached a white landmark on the robot. This landmark is situated in the x-axis of the robot reference system as shown in the robot configuration in Figure 6. The estimation of the robot orientation is based on the pixel coordinates of the landmark centroid. These coordinates are reached using webcam data and the pre-processing method described above. Figure 7 shows a binary image with robot and landmark centroids. The orientation angle is calculated beyond of the robot pixels coordinates as shown in Figure 8. So, by a simple difference between the robot and the landmark pixel coordinates, we obtain the orientation variation along the x-axis and the y-axis as described in the following equation.

Fig. 4. Reference Image.

Fig. 5. Binary reference image with four landmarks centroids.

Fig. 6. Real image showing the robot and the landmark.

Fig. 7. Binary image showing the robot and the landmark centroids. Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

(6)

N째 2

2009

ve on the hardware environment described by a Pentium IV PC with a frequency 3GHz, the image acquisition rate is about 5 seconds and the average processing time is about 1:44 seconds. The problem is that by using this time processing we ensure real time localization only for low velocities. So we search for decreasing this processing time. As a possible solution we apply a cropping rectangle to a binary input image in order to reduce the number of the treated pixels. The architecture of the proposed algorithm is presented in Figure 9. In the first step, we have the first binary image with the default webcam image size 640* 480 pixels (see Figure 10a). On this image we calculate the robot position in pixels coordinates as shown in step three, next we define a crop rectangle with the following form: Croprectangle = [Xmin Y min Width Height]

Fig. 8. Robot orientation according to the reference system. 3.3. Optimised pre-processing method By applying the segmentation method described abo-

Fig. 9. Bloc diagram of the optimised pre-processing method. 18

Articles

The two first parameters depend on the robot coordinates so that for a cropping rectangle with the size 240* 240 pixels (see Figure 10b), Xmin and Y min are presented in equations 7-8.

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

(a) First image: 640*480 pixels

N째 2

2009

y-axis. Then, we used a rule of three to transform the robot pixels coordinates to the real ones. Figure 11 describes the positioning architecture adopted.

(b) Cropped image: 240*240 pixels

Fig. 11. Segmentation and positioning architecture.

4. Experimental Results

Fig. 10. Binary first image and the next cropped images. (7) (8) In fact, we have to note that the choice of the cropping rectangle size is based on experimental test carried out in the mobile robot so that between two following images whatever the wheels velocities, the robot reminded in the new cropped environment. Next (step five), the obtained cropping rectangle is applied on the next acquired image (see Figure 10c) and the robot position is calculated. In step six, we defined the robot displacement beyond the robot position which is in reality the cropping rectangle centre. The step seven is used to remind the robot centroid coordinates in the reference image size defined by 640*480 pixels. The proposed optimised pre-processing method ensured an average processing time equal to 0:48 seconds and which corresponds to 70% time decrease. 3.4. Transformation from image coordinates to real coordinates We obtain the real image coordinates by a simple difference between the pixel coordinates either of the robot or of the reference system. The resulting coordinates are multiplied by a constant coefficient . This coefficient is calculated on the basis of both the real and the pixels distance between the landmarks along the x-axis and the

The proposed method in the previous sections has been implemented and tested on the mobile robot Khepera II. This robot is designed at the Swiss Federal Institute of Technology in Lausanne. It is widely used for research and teaching purposes because it allows real-world testing of algorithms developed in simulation. This Khepera II is circular of 55 mm in diameter and 30 mm in height. Its weight is only 70 g and its small size allows experiments to be performed in small work areas (see Figure 12). It has two driving wheels, which can be independently controlled. The wheel diameter is 15.2 mm. Two incremental encoders placed on each motor axis gives 12 pulses per mm. The distance between the two wheels (L) is 53 mm. This robot has a series of functions, which ensure its control either by using Matlab environment, or Lab-view environment. Two precautions to be taken during the robot command in real time. The first one is the communication speed between the robot and the host computer. The second constraint is to make sure that the wheel speed has already responded to the first instruction given to it before we could send the next one. In this work, the experimental environment used is illustrated in Figure 13. The system studied consists of a robot controlled via a serial communication from the PC. The terminal configuration for the host computer PC must be set to 57600 Bauds, 8 bit, 1 start bit, 2 stop bit and no parity. The robot is equipped with two encoders placed on the left and right wheels and height infrared sensors. The host computer PC executes the task of calculating the optimised trajectory using the (FLC), and determining the robot relative position using odometer measurements. Besides PC is reserved to estimate the position of the robot using visual system. The communication with the webcam is ensured via USB port, with a simple protocol to acquire data. Experimental tests are carried out with the Matlab environment. We have also developed an interface ensuring the real time robot localization with both encoders and webcam data. The interface presented in Figure 14 resumed all steps we need to control the robot. First, we search for the robot initial position and orientation by using an external sensor defined by the webcam as the encoders are unable to provide these parameters. Next we choose a virtual target to be attempted by the robot and the optimiArticles

Journal of Automation, Mobile Robotics & Intelligent Systems

zation technique to be adopted so that the fuzzy controller calculates the wheels velocities. Finally, we started the experimentation by applying these velocities to the robot and calculating the vehicle current position with encoders and webcam data.

VOLUME 3,

N째 2

2009

4.1. Experiment 1 The purpose of the first experiment is to show the application of the segmentation method in the localization of the mobile robot with the webcam data compared to the measurement of odometers. Figures 15 and 16 show respectively the original input images and the segmented images with robot centroid's in image coordinates while the robot had to attempt a target defined by the coordinates (x = 400 mm; y = 50 mm). Therefore, Figure 17 shows an other target defined by the coordinates (x = 50 mm; y = 400 mm).

Fig. 12. Khepera II robot.

Fig. 15. 17 Frames of the original input image. In this last experiment, there exists an approximately 2.8 cm deflection along the x-axis and 0.3 cm deflection along the y-axis but the encoder measurements are erroneous as shown in Figure 17. However using webcam data, the robot position and orientation become very similar to the real one. In fact, the deflection along the x-axis is very important because of the velocities generated by the fuzzy controller and applied to the mobile base are very important where the difference between initial angle and target angle is important. Fig. 13. Experimental environment overview.

Fig. 14. Robot localization using retiming point.

Fig. 17. Robot localization using encoder measurements.

Fig. 16. Different robot positions in image coordinates (frames must be read from the left to the right). 20

Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

N째 2

2009

4.2. Experiment 2 The purpose of the second experiment is to show the real time localization through both encoders and webcam data. Results of this experiment are presented in Figures 18, 19 and 21. In these experiments, the robot have to attempt targets defined respectively by (x = 100 mm; y = 100 mm), (x = 400 mm; y = 100 mm) and (x = 300 mm; y = 300 mm). We notice that the error between the two measures and the deflection along x-axis and y-axis increased while both the distance and the angle moved increased. In the long trajectory experiment described by Figure 19 we have realized four real positioning robot

measures (M1, M2, M3 and M4) using a ruler as shown in Figure 20. The obtained values proved that the absolute localization(visual date) is better than the odometry localization and that the webcam data are much closed to the real measurements. The results of this comparison are summarized in Table 2. This result is due to the incremental movement of the encoders and the wheels slippage problem. So by adopting an absolute localization, the robot trajectory becomes much closed to the real one. But, to further improve the suggested method we used retiming point during the robot trajectory in order to correct the encoder measu-

(a) Robot localization

(b) Robot velocities

Fig. 18. Robot localization and velocities using both enco-ders and webcam measurement during short trajectory. (a) Robot localization

(b) Robot velocities

Fig. 19. Robot localization and velocities using both enco-der and webcam measurements during long trajectory.

Fig. 20. Different robot real positions measured with a ruler. Table 2. Performance evaluating table.

Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

(a) Robot localization

N° 2

2009

(b) Robot velocities

Fig. 21. Robot localization and velocities using both encoder and webcam measurements during a non-linear trajectory. rements by the visual system measures. 4.3. Experiment 3 The purpose of the third experiment is to show the real time localization using the retiming visual control. Results of this experiment are presented in Figure 22. In this experiment, the robot has to attempt the target defined by (x = 300 mm; y = 300 mm) from the initial position defined by the coordinates (x = 50 mm; y = 50 mm).

before we could send the next one. So, the number of times we have to stop the robot depends on the command number sent needed to complete the mission(to reach the target). In fact, this constraint does not effect the estimated position neither with encoders nor with webcam. But to avoid this problem we suggested using the visual system only at few retiming point so that we correct the robot localization and we decrease the time processing. Besides and to further improve the time processing we suggested to use cropping rectangle with a dynamic size.

AUTHORS Mohamed Jallouli*, Lobna Amouri, Nabil Derbel Research unit on Intelligent Control, design and Optimisation of complex Systems (ICOS) University of Sfax, Sfax Engineering School, BP W, 3038 Sfax, Tunisia. Phone: (216-74) 274.088, Fax. (216-74) 275.595. E-mails: mohamed.jallouli@enis.rnu.tn, lobnaamourijmail@yahoo.fr. * Corresponding author Fig. 22. Robot localization using retiming point. References We notice that with the new correction the mobile robot attempts the virtual target with a small error. In fact, we consider one retiming point at the last robot position. The calculated position by the visual system is considered as the new robot initial position used in the optimised fuzzy controller to generate the new wheels velocities in order to attempt the initial virtual target chosen at the beginning.

[1]

[2]

[3]

5. Conclusion In this paper, we have proposed a retiming visual controller for the control of a wheeled mobile robot. Gradient method is used to optimise consequences of a Sugeno fuzzy logic optimal controller used to generate velocities to be applied on the mobile base. The experimental results effected in the Khepera II robot demonstrate that each behaviour works correctly and showed the effectiveness of this method in long and non-linear trajectories. However, to ensure the real time localization using webcam data and because of the visual system time processing and time acquisition we have to stop the robot. This last constraint is to make sure that the wheel speed has already responded to the first instruction given to it 22

Articles

[4]

[5]

[6] [7]

Thomas D. L., Optimal Fusion of Sensors, Department of Automation Technical University of Denmark, PHDEssay, 1998, pp. 1-9. Green D.N., Sasiadek J.Z., and Vukovich G.S., ”Path tracking obstacle avoidance and deal reckoning by an autonomous planetary rover”. In: Proc. IEEE International Conference on Robotics and Automation, San Diego, CA, 1994, pp. 1300-1305. Carelli R., Santos-Victor J., Roberti F., Tosetti S., “Direct visual tracking control of remote cellular robots”. Robotics and Autonomous Systems Journal, vol. 54, July, 2006, pp. 805-814. Escamilla P.J., Neil M., ”Hybrid Kalman Filter-Fuzzy Logic Adaptive Multisensor Data Fusion Architectures”. In: Proc. Decesion and Control Conference, Hawaii USA, December 2003, pp. 5215-5220. Gaone Y., Krakiwsky J., Abousalem M.A., and McLellan J.F., ”Comparison and analysis of centralized, decentralized, and federated filters Navigation”, Journal of the Institute of Navigation, vol. 40, no. 1, 1993, pp. 69-89. Sasiadek J.Z., Hartana P., ”Sensor Data Fusion Using Kalman Filter”. In: Proc. Conf ISIF, 2000, pp. 19-25. Kobayashi F., Arai F., Fukuda T., Shimojima K., Onoda M.

Journal of Automation, Mobile Robotics & Intelligent Systems

[8]

[9]

[10]

[11]

[12]

[13]

[14]

[15]

[16]

VOLUME 3,

N° 2

2009

and Marui N., ”Sensor Fusion System using recurrent fuzzy inference”. Journal of Intelligent and Robotic Systems, vol. 23, 1998, 201-216. Taeseok J., Soomin P., Jang Myung L., ”A Study on Position Determination For Mobile Robot Navigation in an Indoor Environment”. In: Proc. International Symposium on Computational Intelligence in Robotics and Auth th tomation IEEE Conf, Kobe, Japan, 16 -20 July, 2003. Myong Ho K., Sang Cheol L., Kwae Hi L., Self Localization of Mobile Robot With Single Camera in Corridor Environment”. In: Proc. Conf ISIE, vol. 3, June 2001, pp. 1619-1623. Freda L., Oriolo G., “Vision-based interception of a moving target with a non-holonomic mobile robot”. Robotics and Autonomous Systems Journal, vol. 55, February 2007, pp. 419-432. Surmann H., Husser J., Peters L., ”A Fuzzy system for indoor mobile robot navigation”. In: Proc. IEEE international Conference on Fuzzy Systems, Yokohama, Japan, vol.1, 1995, pp. 83-88. Ti-Chung L., Kai-Tai S., Ching-Hung L., Ching-Cheng T., ”Tracking control of Mobile Robots Using Saturation Feedback Controller”. In: Proc. IEEE International Conference on Robotics and Automation, Detroit, Michigan, May 1999, pp. 2639-2644. Prahlad V., Ooi C. M., Xiao P., Tong H. L., ”Fuzzy Behavior-Based Control of Mobile Robots”. In: Proc. IEEE Transactions on Fuzzy Systems, vol. 12, no.4, August 2004, pp. 559-563. Cupertino F., Giordano V., Naso D., .Delfine L., ”Fuzzy Control of a Mobile Robot”. In: Proc. IEEE Robotics and Automation Magazine, December 2006, pp. 74-81. Hung-Ching L., Chih-Ying C., ”The Implementation of Fuzzy-Based Path Planning for Car-Like Mobile Robot”. In: Proc. International Conference on MEMS, NANO and Smart Systems (ICMENS'05), 2005 IEEE, pp. 467-472. Sung-Yug C., Jang Myung L., Chung Kun S., Hyek Hwan C., ”The detection of Lanes and Obstacles in Real Time Using Optimal Moving Window”. JSME International Journal Series - C, vol. 44, no. 2, June 2001, pp. 567-578.

Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 2

2009

ON PATH PLANNING FOR MOBILE ROBOTS: INTRODUCING THE MEREOLOGICAL POTENTIAL FIELD METHOD IN THE FRAMEWORK OF MEREOLOGICAL SPATIAL REASONING Received 24th June 2008; accepted 29th September 2008.

Paweł Ośmiałowski

Abstract: Path planning is one of the most vital problems in mobile robotics; it falls into general province of planning, however, due to specificity of the subject of mobile robotics, it has emerged as a discipline per se with its own solutions. Among many methods of probabilistic, geometrical and topological nature, the methodology of potential fields introduced by Krogh (1984) and Khatib (1985), based on physical analogies with gravitational or electromagnetic fields, has emerged. We adhere to this methodology, however, contrary to the practice of building the potential on the basis of Coulomb, or gravitational force fields, we apply the novel idea of building the potential function by means of mereological distance over a juxtaposition of grids of fixed diameter, i.e., over a discrete structure. We describe our implementation of the relevant mereological functors in the Player/Stage system as SQL predicates accessible in Player/Stage cooperating with PostgreSQL database. We present also the results of simulations with mobile robots. Keywords: mobile robotics, path planning, rough mereology, potential field, Player/Stage system.

1. Introduction Tasks of mobile robotics like path planning, navigation, localization, require for their effective performing adequate representation of the robot environment including obstacles, landmarks, beacons, other robots, along with other static or dynamic features as well as an effective reasoning scheme. Mobile robotics avails itself to this aim with many ideas and methods known in relevant areas of Computer Science and Artificial Intelligence and it does make use of graph methods and algorithms like A* in search or planning [1], topological (graph) representations in map building [1], [25], Markov models in navigation and localization [25] etc. In this work, we propose a new method of constructing a potential field [15] by which to delineate a path for a mobile robot. Our method is based on reasoning scheme for spatial objects developed in the framework of rough mereology [18]. Our potential field inherits the basic property of potential fields, see e.g., [1], i.e., the density of the field does increase in the direction to the target, reaching at the target the maximum value (which in classical cases is infinity). We leave the technical exposition to the following consecutive sections in which we discuss: basic principles of spatial reasoning by mereological predicates, the idea of a potential field along with details of our construction, the description of the Player/Stage system along with SQL predicates implementing basic relevant 24

Articles

mereological relations [19], and finally, results of simulations with mobile robots.

2. Mereological spatial reasoning It is well-known [23] that relations among geometrical objects like solids in 3D or planar figures, are best expressed in the language of parts rather than in the language of set theory: one accepts the statement: "the circle is a part of the closed disc" but one would reject the statement: "the circle is an element of the closed disc"; in the first statement one would accept "a subset" in place of "a part" and indeed, the relation of a subset is a particular instance of the relation of part. The part relation is a basic relation of mereology - the theory of sets/concepts [16] proposed by S. Lesniewski (1916) and adopted to formalize elementary geometry of solids [23]. Due to this fact, one may expect that mereology based constructions will prove useful in mobile robotics tasks. We offer a concise introduction to this area. 2.1. Mereology The part relation between objects e.g., solids is a relation p such that (1) it is not true that p(A, A)for any A, (2) if p(A, B) and p(B, C) then p(A, C). Example: the relation of strict containment Ì is a part relation. The relation of an element (ingredient) el is defined as: el(A, B) if and only if p(A, B) or A = B. Example: the relation of non-strict containment Í is an element relation induced by the part relation Ì. The part, or element, relations give a strict hierarchy of objects consisting of "smaller" objects and so on. However, it was desirable to measure a degree of closeness between e.g. solids not necessarily in the part relation. This was proposed in the frame of rough mereology [18]. 2.2. Rough mereology The basic relation/predicate of rough mereology [18], is a rough inclusion μ(A, B, r), where r Î [0, 1], which means: "A is a part of B to a degree of r". The predicate μ captures our basic intuitions about the nature of containment to a degree and accordingly we impose the following restrictions on it [20]: (1) μ(A, B, 1) if and only if el(A, B) where el is a chosen element relation of a mereology. (2) if μ(A, B, 1) then [if μ(C, A, r) then μ(C, B, r)] for each C. (3) if μ(A, B, r) and s < r then μ(A, B, s).

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

As far as solids or planar figures are concerned, we apply the rough inclusion μ defined as μ(A, B, r) if and only if ||AÇB|| ³ r, ||A||

N° 2

2009

ctor T(X,Y) of two individual names; the statement Z is T(X,Y) reads as Z is between X and Y: Z is T(X,Y) Û

(1)

where ||A|| is the Euclidean volume of A in 3D case, or area in 2D case. 2.3. Qualitative spatial reasoning: Basic geometric predicates induced by μ Qualitative Reasoning aims at studying concepts and calculi on them that arise often at early stages of problem analysis when one is refraining from qualitative or metric details, cf., [5] as such it has close relations to the design, cf., [3] as well as planning stages, cf., [7] of the model synthesis process. Classical formal approaches to spatial reasoning, i.e., to representing spatial entities (points, surfaces, solids) and their features (dimensionality, shape, connectedness degree) rely on Geometry or Topology, i.e., on formal theories whose models are spaces (universes) constructed as sets of points; contrary to this approach, qualitative reasoning about space often exploits pieces of space (regions, boundaries, walls, membranes) and argues in terms of relations abstracted from a commonsense perception (like connected, discrete from, adjacent, intersecting). In this approach, points appear as ideal objects (e.g., ultrafilters of regions/solids [23]). Mereological ideas have been early applied toward axiomatization of geometry of solids, cf., [14], [23]. Mereological theories dominant nowadays come from ideas proposed independently by Stanislaw Lesniewski and Alfred North Whitehead. Mereological theory of Lesniewski is based on the notion of a part (proper) cf., [16]. Mereology based on connection gave rise to spatial calculi based on topological notions derived there from (mereotopology), cf., [5], [6], [8]. Predicates μr may be regarded as weak metrics also in the context of geometry. From this point of view, we may apply μ in order to define basic notions of rough mereological geometry. In the language of this geometry, we may approximately describe and approach geometry of objects; a usage for this geometry may be found, e.g., in navigation and control tasks of mobile robotics [1], [12]. It is well-known, see [24], [2] that the geometry of Euclidean spaces may be based on some postulates about the basic notions of a point and the ternary equi-distance functor. In [24], postulates for Euclidean geometry over a real-closed field were given based on the functor of betweenness and the quaternary equi-distance functor. Similarly, in [2], a set of postulates aimed at rendering general geometric features of geometry of finite-dimensional spaces over reals has been discussed, the primitive notion there being that of nearness. We first introduce a notion of distance kr in our rough mereological universe by letting kr(X, Y) Û r = min{max u, max w : X is μuY Ù Y is μw X}. We now introduce the notion of betweenness as a fun-

for all W kr(Z, W) Ù ks(X, W) Ù kt(Y, W) Þ s £ r £ t Ú t £ r £ s. Thus, Z is T(X,Y) holds when the rough mereological distance k between Z and any W is in the non-oriented interval (i.e. between) [distance of X to W, distance of Y to W] for any W. One checks that T satisfies the axioms of Tarski [24] for betweenness. Proposition 1. The following properties hold, see [19]: 1. Z is T(X, X) Û Z = X (identity); 2. Y is T(X, U) Ù Z is T(Y, U) Þ Y is T(X, Z) (transitivity); 3. Y is T(X, Z) Ù Y is T(X, U) Ù X ¹ Y Þ Z is T(X, U) Ú U is T(X, Z) (connectivity). 2.4. Nearness We may also apply k to define in our context the functor N of nearness proposed in van Bentham [2]: Z is N(X, Y) Û (kr(Z, X) Ù ks(X, Y) Þ s < r). Here, nearness means that Z is closer to X than to Y (recall that rough mereological distance is defined in an opposite way: the smaller r, the greater distance). Then the following hold, i.e., N does satisfy all axioms for nearness in [2], see [19]. Proposition 2. The following properties hold, see [19]: 1. Z is N(X, Y) Ù Y is N(X, W) Þ Z is N(X, W) (transitivity); 2. Z is N(X, Y) Ù X is N(Y, Z) Þ X is N(Z, Y) (triangle inequality); 3. non(Z is N(X, Z)) (irreflexivity); 4. Z = X ? Z is N(Z, X) (selfishness); 5. Z is N(X, Y) Þ Z is N(X, W) Ú W is N(X, Y) (connectedness). We now may introduce the notion of equi-distance as a functor Eq(X, Y) defined as follows: Z is Eq(X, Y) Û (non(X is N(Z, Y)) Ù non(Y is N(Z, X))). It follows that Proposition 3. Z is Eq(X, Y) Û (for all r (kr(X, Z) Û kr(Y, Z)). We may also define a functor of equi-distance following Tarski [24]: D(X, Y, Z, W) Û (for all r kr(X, Y) Û kr(Z, W)). These functors do clearly satisfy the following, see [2], [24], Proposition 4. The following properties hold, see [19]: Articles

Journal of Automation, Mobile Robotics & Intelligent Systems

1. Z is Eq(X, Y) Ù X is Eq(Y, Z) Þ Y is Eq(Z, X) (triangle equality); 2. Z is T(X, Y) Ù W is Eq(X, Y) Þ D(Z, W, X, W) (circle property); 3. D(X, Y, Y, X) (reflexivity); 4. D(X, Y, Z, Z) Þ X = Y (identity); 5. D(X, Y, Z, U) Ù D(X, Y, V, W) Þ D(Z, U, V, W) (transitivity). One may follow van Bentham's proposal for a betweenness functor defined via the nearness functor as follows: Z is TB(X, Y) Û [for all W (Z is W Ú Z is N(X, W) Ú Z is N(Y, W))].

3. Potential fields methodology: Mereological approach Methodology of potential fields [11], [9] was conceived as a framework in which one could generate smooth trajectories from the start point to a target point by mbile robots as well as manipulators. It is known [10] that the method is capable of falling into local minima (as general hill climbing methods do) or oscillatory behaviour, however, working with a global potential map prevents these phenomena: such is our approach in this work. Classical methodology works with integrable force field given by formulas of Coulomb or Newton which prescribe force at a given point as inversely proportional to the squared distance from the target; in consequence the potential is inversely proportional to the distance from the target. The basic property of the potential is that its density (=force) increases in the direction toward the target. We observe this property in our construction. 3.1. The construction of a potential field: The idea We construct the potential field by a discrete construction. The idea is to fill the free workspace of a robot

VOLUME 3,

N° 2

with squares of fixed size in such a way that the density of the square field (measured e.g. as the number of squares intersecting the disc of a given radius r centred at the target) increases toward the target. To ensure this property, we fix a real number - the field growth step in the interval (0, square edge length), in our case it is 0.01. The collection of squares grows recursively with the distance from the target by adding to a given square in the (k + 1)-th step all squares obtained from it by translating it by k ´ field growth step (with respect to Euclidean distance) in basic eight directions: N, S, W, E, NE, NW, SE, SW (in the implementation of this idea, the floodfill algorithm with queue has been used, see the next section). Once the square field is constructed, the path for a robot from a given starting point toward the target is searched for. The idea of this search is in finding a sequence of waypoints which delineate the path to the target. Waypoints are found recursively as centroids of unions of squares mereologically closest to the square of the recently found waypoint. In determining mereological distance, the formula (1) is applied. Parameters of this procedure are: the size of a square (in this work, our robots have been built on the basis of Roomba robot and accordingly, squares we chose to be of edge length equal to 1.5 Roomba (is the trademark of iRobot Inc.) diameter which does ensure safe transition from one square to another close one) and the field growth step (we set it to 0.01 which is small value in comparison to the square size).

4. The mereological path planner The path planner we designed accepts target point coordinates and provides list of waypoints from given robot position to the goal. To do its job it needs a map of static obstacles that a robot should avoid while appro-

Fig. 1. Map of our artificial world edited by the uDig appli-cation (created and maintained by Refractions Research). The map consists of number of layers whose can be edited individually; on the figure we can see how solid obstacles are situated within obstacles layer. 26

Articles

2009

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

aching target point. A robot and a target should both lay within the area delimited by surrounding static obstacles that form borders of robot workspace. There can be other static obstacles within the area, all marked on the provided map. After the path is proposed a robot is lead through the path until it reaches given target. If a robot cannot move towards goal position for some longer time (e.g. it keeps on hitting other robot reaching its target or any other unknown non-static obstacle), new path is proposed. We tested our planner device running simulations in which we have had a model of Roomba robot traveling inside artificial workspace. Real Roomba robots are round and therefore easy to model, however they do not provide many useful sensor devices (except bumpers which we were using to implement lower-level reaction for hitting unexpected obstacles). Also odometry of Roomba robots is unreliable [26] hence we assume that simulated robots are equipped with a global positioning system. Right after the goal position is given, our planner builds mereological potential field filled with squared areas each of the same size. The field is delimited by workspace borders. Only space free of obstacles is filled. To compute a value of potential field in a place we are taking mereological feature of one object being a part of another to a degree where our objects are squared areas that fill the potential field. Near the goal any two squared areas are parts of each other to a higher degree and this value goes low as the distance to the goal increases. It can happen that for bigger workspace, areas too far from the goal are not filled as the potential is limited to values from 0 to 1, where value 0 means that two squares are not part of each other (maximal mereological distance between two areas) while 1 means that two areas are part of each other to a maximal degree (minimal mereological distance between two areas). As a result our potential field is dense with squared areas close to the target and it gets loose far from it.

N째 2

2009

The algorithm of filling the potential field with squared areas is following. SQUARE_FILL_ALGORITHM Structure: a queue Q 1. Add to the queue Q, x and y coordinates of a given goal together with 0 as current distance from current squared area to the next neighbouring area (so they will be part of each other to the maximal degree). Also put clockwise as current direction of exploration. These are initial values. 2. Spin in the main loop until there are no more elements in the queue Q: 2.1. Extract x, y, current distance and current direction of exploration from the beginning of queue Q. 2.2. Check if there is any other squared area already present in potential field to which the distance from current x and y coordinates is equal or shorter than current distance. If so, skip taken element and run new main loop turn. 2.3. Form new squared area with current x and y as the coordinates of the centroid of this new area. Check if there are any common part with any static obstacle within this new squared area. If so, skip taken element and run new main loop turn. 2.4. Add new squared area to the potential field. 2.5. Increase current distance by 0.01. 2.6. Add eight neighbour areas to the queue Q (for each area add these data: x and y coordinates, current distance and direction of exploration opposite to current); if direction is clockwise neighbours are: left, left-up, up, right-up, right, rightdown, down, left-down; if direction is anti-clockwise neighbours are: left-down, down, rightdown, right, right-up, up, left-up, left. 2.7. Run new main loop turn.

Fig. 2. Obstacles layer together with potential field layer (potential field generated for given goal is stored as ano-ther map layer). Observe increasing density towards the goal. Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

N째 2

2009

clockwise direction: left-down, down, right-down, right, right-up, up, left-up, left. Newly added element (with thicker lines on the figure) was left-down to element taken from the begin-ning of the queue Q. Note that distance between the two elements is bigger comparing to previous main loop turn. Fig. 3. Initial step of global potential field construction algorithm. New square area with the goal in the middle (denoted as solid dot) was added as the first element of the potential field.

Fig. 4. First run of the main loop of the potential field construction algorithm. Element on the left was added within the distance 0.01 from the initial square. In this main loop turn, elements are added using clockwise direction: left, left-up, up, right-up, right, right-down, down, left-down.

Fig. 7. Second run of the main loop of the potential field construction algorithm is now completed. Eight more squares were added to the potential field (and queue Q) around first square taken (and removed) from the queue (square with thicker lines on the figure). The algorithm of searching for a path within given potential field is following: PATH_SEARCH_ALGORITHM

Fig. 5. First run of the main loop of the potential field construction algorithm is now completed. Eight squares were added to the potential field around initial square. These squares were also added to the queue Q so they will be operated in next main loop turns.

Fig. 6. First element from the beginning of the queue Q was taken - it is the square to the left of the initial area. In this main loop turn, elements are added using anti-

1. Check if there is a field coverage in the current robot position. Also check if the robot is not at the goal which means no path needs to be planned. Choose a start square from the potential field: the nearest squared area from areas that have any common part with the robot. The centroid of this area will be the first waypoint. 2. Spin in the main loop until goal is found: 2.1. From the set of squared areas that have any common part with previously choosen squared area choose the one which is the part of this area to the bigest degree (the smallest mereological distance). The centroid of this area will be the next waypoint. 2.2. Run new main loop turn. A robot should follow the path proposed by planner by going from one area centroid to another until the goal is reached.

Fig. 8. Stage simulator in use - two iRobot Roomba robots inside of simulated world waiting for a goal to be set. 28

Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

5. Implementation in Player/Stage robotics framework Player/Stage is an Open-Source software framework designed for many UNIX-compatible platforms, widely used in robotics laboratories [17]. Main two parts are Player - message passing server (with bunch of drivers for many robotics devices, extendable by plugins) and Stage - a plugin for Players bunch of drivers which simulates existence of real robotics devices that operate in simulated 2D world. Player/Stage offers client-server architecture. Many clients can connect to one Player server, where clients are programs (robot controllers) written by a roboticist who can use Player client-side API. Player itself uses drivers to communicate with devices. In this activity it does not make distinction between real and simulated hardware. It gives roboticist means for testing programmed robot controller in both real and simulated world. Among all Player drivers that communicate with devices (real or simulated), there are drivers not intended for controlling hardware and instead they offer many facilities for sensor data manipulation. For example: camera image compression, retro-reflective detection of cylindrical markers in laser scans, path planning. One of such drivers widely used during our experimentations is the PostGIS driver. It connects to PostgreSQL database [22] in order to obtain and/or update stored vector map layers. PostGIS itself is an extension to the PostgreSQL object-relational database system which allows GIS (Geographics Information Systems) objects to be stored in the database [21]. It also offers new SQL functions for spatial reasoning. Maps which to be stored in SQL database can be created and edited by graphical tools like uDig or by C/C++ programs written using GEOS library of GIS functions. PostGIS, uDig and GEOS library are projects maintained by Refractions Research. A map can have many named layers, for each layer a table in SQL database is created. We can assume that layer named obstacles is full of objects that a robot cannot walk through. Other layers can be created and we are using one of such layers to store potential field data (squared areas). A roboticist can write a robot controller using Player

N째 2

2009

client-side API, which obtains information about current situation through the vectormap interface. Additionally, to write such a program, PostgreSQL client-side API can be used to open direct connection to the database server. A robot controller does not need to be implemented as the client-side program. Other way is to write a C++ code (a plugin) which will act as a new driver for the Players bunch of drivers. Player itself already contains one such driver called wavefront. This driver plays two roles: it acts as a simple path planner and it can controll a robot to walk through the path. It can use another built-in driver called vfh which implements Vector Field Histogram algorithm for obstacle avoidance [4]. Our mereological path planner is implemented as a plugin driver that can replace wavefront driver. However, it is not intended to use vfh driver for obstacle avoidance as the Vector Field Histogram algorithm needs any ranger sensor (laser, sonar, infrared) while Roomba robot default configuration does not have any such device. Instead, robot controller part of our driver monitors how long does it take to achieve next waypoint and if it is too long, it asks planner to replan the path from current robot position using already computed potential field. Additionally, we have added to our driver low-level behaviour that monitors state of bumpers and whenever bumpers are closed, the robot is going back-left to a new position from which the path is replanned. As on the client-side, server-side drivers can use vectormap interface to obtain required map layer. Also direct connection to PostgreSQL database server can be opened. In our planner driver we have used ECPG API provided by PostgreSQL which enables to put SQL queries directly into the C/C++ code. To make our SQL queries more robust, we have stored our mereogeometry SQL functions on PostgreSQL server together with map database. These functions can be called using connection with database managed by ECPG infrastructure. Our mereological functions are processed on SQL database server side, results are sent back to the calling program (which means to our planner). This gives our planner ability to perform spatial reasoning based on rough mereology. We have created our mereogeometry SQL predicates [13]. Rough mereological distance is defined as such:

CREATE FUNCTION meredist(object1 geometry, object2 geometry) RETURNS DOUBLE PRECISION AS $$ SELECT min(degrees.degree) FROM ((SELECT ST_Area(ST_Intersection(extent($1), extent($2))) / ST_Area(extent($1)) AS degree) UNION (SELECT ST_Area(ST_Intersection(extent($1), extent($2))) / ST_Area(extent($2)) AS degree)) AS degrees; $$ LANGUAGE SQL STABLE; Having mereological distance function we can derive nearness predicate: Articles

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 3,

N째 2

2009

CREATE FUNCTION merenear(obj geometry, o1 geometry, o2 geometry) RETURNS BOOLEAN AS $$ SELECT meredist($1, $2) > meredist($3, $2) $$ LANGUAGE SQL STABLE; The equi-distance can be derived as such: CREATE FUNCTION mereequ(obj geometry, o1 geometry, o2 geometry) RETURNS BOOLEAN AS $$ SELECT (NOT merenear($1, $2, $3)) AND (NOT merenear($1, $3, $2)); $$ LANGUAGE SQL STABLE; Our implementation of the betweenness predicate makes use of a function that produces an object which is an extent of given two objects: CREATE FUNCTION mereextent(object1 geometry, object2 geometry) RETURNS geometry AS $$ SELECT GeomFromWKB(AsBinary(extent(objects.geom))) FROM ((SELECT $1 AS geom) UNION (SELECT $2 AS geom)) AS objects; $$ LANGUAGE SQL STABLE; The betweenness predicate is defined as such: CREATE FUNCTION merebetb(obj geometry, o1 geometry, o2 geometry) RETURNS BOOLEAN AS $$ SELECT meredist($1, $2) = 1 OR meredist($1, $3) = 1 OR (meredist($1, $2) > 0 AND meredist($1, $3) > 0 AND meredist(mereextent($2, $3), mereextent(mereextent($1, $2), $3)) = 1); $$ LANGUAGE SQL STABLE; Using the betweenness predicate we can check if three objects form a pattern: CREATE FUNCTION merepattern(object1 geometry, object2 geometry, object3 geometry) RETURNS BOOLEAN AS $$ SELECT merebetb($3, $2, $1) OR merebetb($1, $3, $2) OR merebetb($2, $1, $3); $$ LANGUAGE SQL STABLE; Also having pattern predicate we can check if four objects form a line: CREATE FUNCTION mereisline4(obj1 geometry, obj2 geometry, obj3 geometry, obj4 geometry) RETURNS BOOLEAN AS $$ SELECT merepattern($1, $2, $3) AND merepattern($2, $3, $4); $$ LANGUAGE SQL STABLE; As we can realise, the path from current robot position to the goal is built from squared areas that form mereological line (as described by predicates above). We derived SQL aggregate function that can check if a given set of areas form a mereological line. This consists of one state function, one final function and one aggregate definition: 30

Articles

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 3,

N째 2

2009

CREATE FUNCTION mereislinestate(statearray geometry[4], inputdata geometry) RETURNS geometry[4] AS $$ SELECT ARRAY[$1[2], $1[3], $2, result.object] FROM (SELECT CASE WHEN $1[4] IS NOT NULL THEN $1[4] WHEN $1[3] IS NULL THEN NULL WHEN ($1[2] IS NULL) AND (meredist($1[3], $2) > 0) THEN NULL WHEN ($1[2] IS NULL) AND (meredist($1[3], $2) = 0) THEN $2 WHEN ($1[1] IS NULL) AND merepattern($1[2], $1[3], $2) THEN NULL WHEN ($1[1] IS NULL) AND (NOT merepattern($1[2], $1[3], $2)) THEN $2 WHEN merepattern($1[1], $1[2], $1[3]) AND merepattern($1[2], $1[3], $2) THEN NULL ELSE $2 END AS object) AS result; $$ LANGUAGE SQL STABLE;

Fig. 9. Playernav in use - Roomba robot waits for the orders. CREATE FUNCTION mereislinefinal(statearray geometry[4]) RETURNS BOOLEAN AS $$ SELECT ($1[4] IS NULL) AND ($1[3] IS NOT NULL) AND ($1[2] IS NOT NULL); $$ LANGUAGE SQL STABLE; CREATE AGGREGATE mereisline ( SFUNC = mereislinestate, BASETYPE = geometry, STYPE = geometry[], FINALFUNC = mereislinefinal, INITCOND = '{}' ); One of client-side programs that come together with Player server is playernav. It shows a map of robot environment (with marked current robot position). We can in-dicate on the map goal position where robot should go. The playernav sends the goal to planner device working on given Player server instance. Then it asks the planner for current path which will be marked on the map.

Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 2

2009

Fig. 10. Playernav in use - Roomba robot follows path to the target.

Fig. 11. Show trails is a nice option in Stage which can be used to track robot trajectory. Here we can see how two Roomba robots walked through planned paths to their targets.

6. Comparison with Player's built-in path planning device As it was stated before, Player already provides its own path planning device called wavefront. This mechanism provides correct paths to given targets. Our method is implemented as a plugin for Player, which acts as a direct replacement for wavefront and provides correct paths too. It is worth to mention that our method gives possibility of observing intermediate stage of its routine: computed potential field used by planner can be viewed using uDig software. This gives possibility of additional tweaking of the algorithm. Each planner in certain circumstances must perform replanning. To do this wavefront must repeat whole planning routine. Using our method, only second stage of planning routine is done during replannig as potential field is computed only once (unless the database is updated with new obstacles). Searching for a path within already computed potential field is computationaly cheap as it is limited to database lookup operations (therefore speed of database communication is critical if our method is intented to be working fast).

AUTHOR Paweł Ośmiałowski - Polish-Japanese Institute of Information Technology, Chair of Intelligent Robotic Systems, Koszykowa str. 86, 02-008 Warszawa, Poland. E-mail: newchief@king.net.pl.

References [1] [2] [3] [4]

[5]

7. Conclusions Results of simulations show that this method gives satisfactory results. Further research is aimed at planning paths for teams of robots as well as for planning paths for prescribed robot formations.

Articles

[6]

Arkin R. C., Behavior Based Robotics, MIT Press: Cambridge MA, 1998. van Bentham J., The Logic of Time, Reidel: Dordrecht, 1983. Booch G., Object-Oriented Analysis and Design with Applications, Addison-Wesley Publ., Menlo Park, 1994. Borenstein J., Koren Y., “The Vector Field Histogram Fast Obstacle Avoidance for Mobile Robots”, IEEE Journal of Robotics and Automation, vol. 7, no. 3, June 1991, pp. 278-288. Cohn A. G., “Calculi for qualitative spatial reasoning”, in: J. Calmet, J. A. Campbell, J. Pfalzgraf (eds.), Artificial Intelligence and Symbolic Mathematical Computation, Lecture Notes in Computer Science, vol. 1138, Springer Verlag: Berlin, 1996, pp. 124-143. Cui Z., Cohn A. G., Randell D. A., “Qualitative and topological relationships”, in: Advances in Spatial Databases, Lecture Notes in Computer Science, vol. 692, Springer Verlag: Berlin, 1993, pp. 296-315.

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

[7]

[8]

[9]

[10]

[11]

[12]

[13] [14] [15] [16] [17]

[18]

[19]

[20]

[21] [22] [23]

[24]

Glasgow J., “A formalism for model-based spatial planning”, in: A. U. Frank, W. Kuhn (eds.), Spatial Information theory - A Theoretical Basis for GIS, Lecture Notes in Computer Science, vol. 988, Springer Verlag, Berlin, 1995, pp. 501-518. Gotts N. M., Cohn A. G., “A mereological approach to representing spatial vagueness”. In: Working papers, the Ninth International Workshop on Qualitative Reasoning, QR'95, 1995. Khatib O., “Real-time obstacle avoidance for manipulators and mobile robots”. In: Proceedings IEEE Intern. Conf. on Robotics and Automation, St. Louis MO, pp. 500-505. Koren Y., Borenstein J., “Potential field methods and their inherent limitations for mobile robot navigation”. In: Proceedings of the IEEE International Conference on Robotics and Automation, Sacramento CA, 1991, pp. 1398-1404. Krogh B., A generalized potential field approach to obstacle avoidance contro”, SME-I Technical paper MS84484, Society of Manufacturing Engineers, Dearborn MI, 1984. Kuipers B., Qualitative Reasoning: Modeling and Simulation with Incomplete Knowledge, MIT Press: Cambridge MA, 1994. Ladanyi H., SQL Unleashed, Sams Publishing, 1997. de Laguna T., Point, line, surface as sets of solids, J. Philosophy, vol. 19, 1922, 449461. Latombe J.-C., Robot Motion Planning, Kluwer: Boston, 1991. Lesniewski S., “On the foundations of mathematics”, Topoi 2, 1982, pp. 752. P. Osmialowski, “Player and Stage at PJIIT Robotics Laboratory”, Journal of Automation, Mobile Robotics and Intelligent Systems, no. 2, 2007, pp. 21-28. Polkowski L., “An Approach to Granulation of Knowledge and Granular Computing Based on Rough Mereology: A Survey”, in: V. Kreinovich, W. Pedrycz, A. Skowron (eds.), Handbook of Granular Computing, John Wiley and Sons: New York NY, 2008. Polkowski L., Osmialowski P., “Spatial reasoning with applications to mobile robotics”. In: Motion Planning for Mobile Robots: New Advances, InTech, Vienna, 2008. Polkowski L., Skowron A., “Rough mereology in information systems with applications to qualitative spatial reasoning”, Fundamenta Informaticae, vol. 43, issue 14, 2000, pp. 291-320. Ramsey P., PostGIS Manual. In: postgis.pdf file downloaded from Refractions Research home page, 2008. Stones R., Matthew N., Beginning Databases with PostgreSQL, Wrox Press, 2001. Tarski A., “Les fondements de la géométrie des corps”. In: Ksiega Pamiatkowa I Polskiego Zjazdu Matematycznego (Memorial Book of the Ist Polish Mathematical Congress), a supplement to Annales de la Sociéte Polonaise de Mathématique, Kraków (Cracow), 1929, pp. 29-33. Tarski A., “What is elementary geometry?” In: L. Henkin, P. Suppes, A. Tarski (eds.), The Axiomatic Method with Special Reference to Geometry and Physics, Studies in Logic and Foundations of Mathematics, North-Holland: Amsterdam, 1959, pp. 16-29.

[25] [26]

N° 2

2009

Thrun S., Burgard W., Fox D., Probabilistic Robotics, MIT Press: Cambridge MA, 2005. Tribelhorn B., Dodds Z., “Evaluating the Roomba: A lowcost, ubiquitous platform for robotics research and education”. In: 2007 IEEE International Conference on Robotics and Automation, ICRA 2007, 10th-14th April 2007, Roma, Italy, pp. 1393-1399.

Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

N째 2

2009

AN ENCODED INFRARED SHEET OF LIGHT NAVIGATIONAL BEACON SYSTEM FOR PRECISE LOCALIZATION OF INDOOR MOBILE ROBOT VEHICLES Received 12th August 2008; accepted 8th December 2008.

James Kurian, P. R. Saseendran Pillai

Abstract: A new localization approach to increase the navigational capabilities and object manipulation of autonomous mobile robots, based on an encoded infrared sheet of light beacon system, which provides position errors smaller than 0.02m is presented in this paper. To achieve this minimal position error, a resolution enhancement technique has been developed by utilising an inbuilt odometric/optical flow sensor information. This system respects strong low cost constraints by using an innovative assembly for the digitally encoded infrared transmitter. For better guidance of mobile robot vehicles, an online traffic signalling capability is also incorporated. Other added features are its less computational complexity and online localization capability all these without any estimation uncertainty. The constructional details, experimental results and computational methodologies of the system are also described. Keywords: infrared beacons, position estimation; robot localization; sheet of light beacons.

1. Introduction Accurate sensing of vehicle position and attitude is a vital requirement in many mobile robot applications. In this modern age the autonomous or semi autonomous robot vehicles find applications in automated inspection systems [1], floor sweepers [2], hazardous environments [3], autonomous truck loading systems [4], agriculture tasks, delivery in establishments like manufacturing plants, office buildings, hospitals [5], etc. and providing services for the elderly [6]. In addition to this, autonomous vehicles are widely utilized in undersea exploration and military surveillance systems [7,8]. Mobile robots are also finding their way into a growing number of homes, providing security, automation [9,10], and even entertainment. In order to navigate to their destination, the robots must have some means of estimating where they are and in which direction they are heading. Information about the location of an inanimate object, for example a cargo pallet, can streamline inventory and enable warehouse automation. A variety of technologies have been developed and used successfully to provide position and attitude information. However, many of these existing positioning systems have inherent limitations in their workspace. These limitations generally fall into two main categories: line-of-sight restrictions and insufficient resolution/precision as they require multiple clear linesof-sight and absolute drift-free measurements. In mobile robot applications, two basic position estimation methods are employed concurrently, viz., the ab34

Articles

solute and relative positioning [11]. Absolute positioning methods usually rely on the use of appropriate exteroceptive (external) sensing techniques, like navigation beacons [12,13], active or passive landmarks [14], map matching [15], or satellite-based navigation [16] signals. Navigation beacons and landmarks normally require costly installations and maintenance, while map-matching methods are usually slower and demand more memory and computational overheads. The satellite-based navigation techniques are used only in outdoor implementations and have poor accuracy, of the order of a few metres. Relative position estimation or dead reckoning is based on proprioceptive (internal) sensing systems like odometry [17], inertial navigation system (INS) [18] or optical flow techniques [19], where the error growth rate of these systems are usually unacceptable. The vehicle performs self-localization by using relative positioning technique, called dead reckoning. For implementing a navigational system many indoor mobile robots use active beacons [13] together with traditional inertial navigation systems employing gyros and accelerometers or position odometric system or both. The latter provides accurate and precise intermediate estimation of position during the path execution. Inertial Navigation System (INS) is complex and expensive and requires more information processing for extracting the required position and attitude information. The localization based on INS uses accelerometers or gyros, where the accelerometer data must be integrated twice to yield the position information, thereby making these sensors extremely sensitive to drift. Though the odometric system is simple, inexpensive and accurate over short distances, it is prone to several sources of errors due to wheel slippage, variations in wheel radius, body deflections, surface roughness and undulations. For better traction most of the mobile robots use rubber tires, which have unevenness in their diameter and these tires compress differently under asymmetric load distribution or load imbalances, causing further position and attitude errors. For the successful navigation and path planning of mobile robots, a well-defined and structured workspace is required. This can provide high-rate of precise positioning and attitude information for reliable estimation of the vehicles' localization and navigation map. For outdoor applications Differential Global Positioning System (DGPS) based localization techniques provide adequate resolution, whereas for indoor use, this resolution is insufficient and moreover the satellite signals may be obstructed, which further aggravate the situation. Substantial research works are going on in the area of simul-

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

taneous localization and map building (SLAM) [20] using various sensing systems, which require more memory and computational overhead for feature extraction. The errors in kinematic and environmental parameters will lead to poor estimation of positions during the path execution and this necessitates the need for frequent absolute localizations. For indoor applications like localization of personnel, products and vehicles in warehouses as well as production environments, where a stable and accurate localization system is necessary, the ultrasonic, infrared, [27] radio frequency [21] and laser techniques [22] are commonly used. The use of ultrasonic sensors [23, 24] is limited to the proximetry because of poor system characteristics like moderate axial resolution, low lateral resolution, and high rate of inaccuracies in measurements resulting from multiple reflections, environmental complexity and the aperture cone. Radio frequency systems are very expensive and are susceptible to reflections from metallic objects. These localization systems, which utilize triangulation or trilateration techniques [25], have high uncertainty in position estimations, incurring extra computational overheads, resulting possibly in slowing down the path execution process of the vehicle. Most of the high-resolution systems are complex and expensive. A cost effective commercially available infrared Beacon System used for indoor robot localization application is the Northstar from Evolution Robotics Inc. [28]. This system requires a reflecting roof for its functioning which is not always feasible in an industrial/ warehouse environment. The reflective characteristics as well as the indoor lighting system may affect its performance. This system suffers from the computational overheads due to the triangulation technique. The limitations of the above diverse ways of positioning systems that are already in use open the scope for further research opportunities for improvement and innovations. Many applications that depend on position measurements could benefit from the development of a new positioning system technology that alleviates these restrictions. The development of a cost effective, accurate and reliable system, utilising an infrared sheet of light, which minimizes position errors during the path execution is presented in this paper. This provides a cost effective position and attitude sensing system designed specifically to face the challenges in a realistic, cluttered indoor environment, such as that of an office building or warehouse. In the proposed approach, a number of beacon transmitters are installed in the well defined and structured workspace as required and all the transmitters provide the estimates in a common reference frame or even universal frame. Two sensor units on the mobile robot read the beacon and process the measurements to determine its position, attitude and traffic signalling information. The real-time identification and correction methods mitigate the impact of localization errors caused by the robot vehicles and the environment. A novel resolution enhancement algorithm suggested in this paper satisfies the requirements for a high-resolution localization system. A prototype system has been built to demonstrate the suggested approach.

N째 2

2009

2. Sheet of light beacon 2.1. Principle of operation The localization systems based on computer vision, range finders or other sensors that do not require a special arrangement of the environment are computationally expensive and not too robust. Most of the infrared, ultrasonic or radio frequency beacons have inherent emission characteristics that may affect the resolution of the measuring system. Hence it is essential to consider a robust, low-cost system for the absolute positioning of mobile robots or other moving objects. This work describes an assembly utilizing an infrared LED source that restricts the spreading of the light intensity distribution confined to a sheet of light.

Fig. 1. The infrared LED of the beacon transmitter mounted on the structural assembly.

Fig. 2. Variation of effective light sheet thickness against the mounting height of the beacon. Sheet of light techniques are utilized in robotics and industrial applications for sensing objects, its shape and size [26]. Here a new approach to produce the sheet of light and an encoding scheme for localization application is described. In order to produce a sheet of light for highresolution localization applications, an innovative assembly as shown in Fig. 1 has developed. The infrared beam is guided through the space between two identical sand blasted parallel metal plates of dimensions 100 mm Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

x 100 mm, kept 2.5 mm apart. These metal plates will be acting as Lambertian scattering surfaces (diffuse reflectors) and their dimensions have effects on the sheet thickness as well as infrared light intensity. A single infrared LED mounted at the centre of the LED housing, as shown, is seen to have a beam angle of around 45 degrees, which can be increased by mounting multiple LEDs. The width of the region of the infrared light sheet where the receiving system can properly read the encoded position is the Effective Light Sheet Thickness (ELST). The variation of effective infrared light sheet thickness against the height (h) of the mounting structure has been studied and the results are shown in Fig. 2, which illustrates a linear increase in light sheet thickness for mounting heights above two metres. 2.2. The beacon transmitter The Digital Infrared Sheet of Light Beacons (DISLiB) constructed using the above assembly are location encoded and are designed around a 16F675 PIC microcontroller as shown in Fig. 3. The system transmits a carrier frequency of 40 kHz, which is pulse width modulated with 12 bit Beacon Identification Number (BIN), one parity bit and appropriate start pulse. The BIN is assigned to each beacon installed in the workspace. The system employs a scaled version of Sony Infrared Remote Control (SIRC) protocol to transmit the data and the protocol structure is shown in Fig. 4. The time taken to transmit a location information is around 6 ms, which may vary slightly as the protocol uses different burst lengths for '1's and '0's. Besides continuously transmitting the encoded position information the microcontroller in the beacon transmitter drives the infrared LED(s) by switching a transistor in series with a current limiting resistor. By interfacing micro-switch inputs to the microcontroller for the configuration of a particular Beacon Identification Number (BIN), one can easily encode different

3. Vehicle localization 3.1. Method of Installation and Working Most of the absolute localization methods using ultrasonic, infrared, radio frequency or laser require multiple known beacons or encoded strips in the vicinity of the robot vehicle as well as rotating/scanning, control and computational units to estimate the position of the system. If multiple localization systems are installed, a sensor fusion algorithm must be used to obtain a better estimate [27]. A prerequisite for a successful map matching or landmark based technique of localization, is an acceptable accuracy in the relative position estimation. By eliminating all the inherent problems and complexities of these existing systems a high-resolution absolute localization is possible with the use of Digital Infrared Sheet of Light Beacons (DISLiB). By properly installing the Digital Infrared Sheet of Light Beacons (DISLiB) at known locations (B) vertically above the track as shown in Fig. 5, an accurate and robust representation of the workspace

Fig. 4. The scheme of the scaled version of SIRC communication protocol. Articles

2009

location information to the beacons without modifying the firmware in each unit. Traffic signalling information like speed limits, sharp turnings etc., can be communicated by adding a few more bits and properly encoding the beacon. A number of beacon transmitters are mounted at various locations to define the environmental structure. Each beacon will send fixed BIN plus traffic signalling bits to the receiver. By establishing an RS 485 network among the beacons and a host computer, the position information in case of restructuring, as well as traffic signalling commands can be modified online. The RS 485 interface is designed using DS75176 transceiver chip and the RS 485_USB bridge is designed around an 18F2550 PIC microcontroller with inbuilt USB support. Thus the system can be made user friendly by incorporating the RS 485 network with the host computer.

Fig. 3. Functional block diagram of the beacon transmitter.

N째 2

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

can be achieved for path planning and object identification of the mobile robot. In a typical indoor structure, the beacons should be mounted at a height (h) of about three metres for covering the entire width of the track and for greater track widths either multiple infrared LEDs or increased mounting heights within the reading threshold of the beacons are preferred. For a systematic implementation of the system, the entire workspace can be divided into various zones and tracks, where each track in the zones is properly labelled for effective functioning. The beacon distributions can be identified based on the systematic errors resulting from the kinematic imperfections of the vehicle and non-systematic errors due to the environment and depending on the resolution requirements.

N째 2

2009

manages the drive and control systems. A wireless link is established to monitor and assist the navigational guidance system of the robot vehicle, which utilizes an RF Programmable System on Chip (PSoC) module [29] from Cypress Semiconductor Corporation - CYWM6935 PAEC. The module can have a range of about 200 metres and operates at 2.4 GHz ISM band. It has inbuilt Direct Sequence Spread Spectrum (DSSS) communication [30] facility with a 64 bit PN code for spreading and dispreading of data. The on chip serial peripheral interface (SPI) can be utilized for configuration and establishing communication with the module. This RF transceiver module is ideal for short-range indoor applications. For the receiver/controller design microchip PIC 18F4550 40-pin microcontroller with inbuilt SPI support is used. The system performance can further be improved by using multiple microcontroller-based designs.

Fig. 6. Block diagram of the beacon receiver and controller. 3.3. The Beacon performance and evaluation a)

Fig. 5. A typical workspace showing the beacon positions (B) and mounting of the same vertically above the tracks Tr1,Tr2 etc. at a height of h metres. 3.2. The Beacon Receiver and Controller During path execution, the position information gathered by the infrared remote control receiver module from the beacon is processed by the microcontroller system of the vehicle that manages its navigation and guidance. As the vehicle crosses the infrared light sheet of thickness d, the microcontroller based navigation system directly captures the location-encoded information (BIN) and the position is updated after retrieving the corresponding absolute position from the database. The receiver takes 6 ms for position decoding and hence at least 12 ms is required for a guaranteed position update while crossing a DISLiB. For a mounting height of about three metres the effective light sheet thickness (d) is around 0.12 m (Fig. 2), and hence the maximum speed of the vehicle has to be limited to a value less than 10 m/s. As the speed of practical indoor vehicles is less than this, it does not cause any problem in field applications. Up to this speed, the resolution of the system remains as the effective light sheet thickness. The functional block diagram of a typical beacon receiver is shown in Fig. 6. The odometric sensors provide the position information to the microcontroller, which

Fig. 7. 3-D surface plots (a) indicates the variation of resolution with respect to speed and reading time (b) the variation of resolution with respect to speed and ELST. Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

This is an absolute localization system for correcting the errors caused by the inbuilt sensory system of the mobile robot vehicle. In situations where frequent correction is required, more numbers of beacons are to be installed. The beacon (DISLiB) performance is associated with various parameters like the speed of the vehicle (during beacon crossing), effective light sheet thickness (ELST) of the system and the reading time, which depends upon the coding scheme of the beacon transmitter. As the vehicle crosses the DISLiB, the system takes a certain number of readings depending on reading time, ELST and the speed of the vehicle. The role of these parameters, which affect the performance, has been studied and a resolution enhancement algorithm has been developed which is explained in section 4. The characteristics are plotted in Fig. 7. These 3-D surface plots show the role of vehicle speed, beacon's reading time and Effective Light Sheet Thickness on the resolution of the system. Fig. 7a indicates the plot of speed and reading time against the resolution with an effective sheet width of 0.12 metres. The discrete variation of the resolution depends on the number of beacon readings, which is a function of speed and reading time. Fig. 7b shows the effect of ELST on the resolution and vehicle speed.

4. Resolution Enhancement During path execution, as the vehicle crosses the DISLiB, the system takes n number of readings depending on the ELST and the speed of the vehicle. For vehicles moving at a speed less than the maximum speed allowed by the system, the resolution could be increased by making use of a resolution enhancement algorithm. Fast moving vehicles have to be slowed down during the localization process for achieving acceptable resolution enhancement. The system generates a lookup table with the count (n), beacon reading (BIN) and the odomertic position information (Pn), as shown in Table 1. Table 1. Lookup table formulated for the execution of the resolution enhancement algorithm.

Under a particular DISLiB the beacon identification number is the same for all the observations. The beacon identification number points to a memory location in the database from where the position information can be retrieved. The position information furnished by the proth th prioceptive sensors corresponding to n/2 or (n+1)/2 position respectively for even or odd values of n can be updated. For an even value of n, Pn/2 can be replaced with 38

Articles

N° 2

2009

an absolute position value from the database pointed by BIN and hence the resolution of the system is enhanced from effective width of light sheet d to d/n. In fact the th algo-rithm replaces the Pn+1 position value with [BIN] + (Pn+1 - Pn/2). The vehicles fitted with two sensors create separate tables and finally updates with the average value. The resolution is effectively improved in the present set up, as it is the product of the speed of the vehicle and reading time, as illustrated below. The enhanced resolution r, which is the ratio of the effective light sheet thickness to the number of readings, can be deduced to: (1) Where sv - speed of the vehicle tb - time taken by the vehicle to cross the light sheet (ELST) tr - time required for one beacon reading INT(tb/tr) - integer value of the ratio (tb/tr). Equation (1) gives the inference that a reduced vehicle speed improves the resolution. When tr = tb, the resolution enhancement algorithm will fail since n=1, and the resolution remains at d. For a vehicle crossing the DISLiB at a speed of around 3 m/sec. with a reading time of 6ms equation (1) computes the resolution to be approximately 0.02 m. A further improvement in resolution can be obtained by reducing the beacon reading time, which in turn is achieved by decreasing the infrared burst lengths. Usually the reading time is constant for a set up so that the resolution of the system varies with the speed of the vehicle.

5. Position and attitude update The kinematics and navigation equations for a threewheeled mobile vehicle with one driving-steering wheel and two fixed rear wheels in-axis is considered for this study. The odometric navigational systems are implemented using three optical incremental encoders. The driving steering wheel (front) is attached with a permanent magnet DC motor with inbuilt encoder, which measures the angular increments and a multi-turn potentiometer for the measurement of the steering angle . The rear wheels are also attached with encoders to estimate the position and attitude of the vehicle. The DISLiB beacon receivers are utilized to update the position and heading of the vehicle by utilizing the update equations for this vehicle geometry. The symbols used in the equations are defined below: - the angle between the light sheet footprint and the line joining between the beacon receiver sensors S1 and S2. - the angle between the light sheet footprint of the ith DISLiB and y axis of the fixed reference frame. - the estimated attitude of the vehicle with respect to the fixed referenceþ - the estimated attitude of the vehicle using the ith DISLiB with respect to the fixed referenceþ - the steering angle with respect to axis of symmetryþ R - the wheel radius of the vehicleþ nL- nR - nF - the encoder incremental pulse counts from the

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

left, right and front wheel encoders respectively. N - the number of pulses per revolution of the encoder. L - the distance between the rotation axis of the front (driver) wheel and the axis of the back wheel. D - the distance between rear wheels. tb - the time required to cross the beacon light sheet. td - the time delay between two sensor outputs. a - half the distance between two beacon sensors.

N° 2

2009

follows [31]: (3) (4)

(5) The distance moved by the wheel's point of contact could be derived by considering the vehicle's front driving steering wheel's incremental pulse count data. The steering rotation is limited to ±400 about the axis of symmetry of the vehicle. The potentiometer sensor attached to the steering system generates a voltage in the range of 0.5 V to 4.5 V for representing the steering angle. This is fed to the analog to digital converter, the output of which can be read by the microcontroller and the corresponding steering can be computed. The update equations for this model are described as follows: (6)

(7) Fig. 8. Kinematic scheme of the three-wheeled mobile robot vehicle and the footprint of the effective light sheet width d. The attitude is the angle between the absolute reference frame OXY and the mobile reference frame PUV. The origin P is attached to the mid point of the axes joining the rear wheels and the sensors S1 and S2. Fig. 8 shows a typical posture of the mobile vehicle with an orientation “ ” and steering angle “ “. Two identical DISLiB sensors S1 and S2 are mounted at the top of the rear wheel axis of the vehicle at a distance of 2a. If the vehicle's axis of symmetry is normal to the sheet of light both the sensors receive the signal simultaneously. From the BIN received, the mounting angle of the corresponding beacon transmitter, can be retrieved from the database. If the vehicle crosses the beacon with a heading angle “ ” (not equal to ) there will be a lag or lead between the received signals, which is a measure of the attitude of the vehicle. The signal waveforms derived from the start pulse is shown in Fig. 8 (inside the circle), in which the time duration tb is the time required to cross the beacon and the lag or lead time td is the time required to cover the distance “c” by the vehicle. The lead or lag time td is a measure of the attitude of the vehicle. The attitude computed by the receiver unit in the vehicle is given by the following expression: (2) The computed value of , which is in equations (5) and (8) is updated with this . For the computation of the position and attitude let us consider the pulse counts from the two independent optical encoders attached to the rear non-driven idler wheels of the vehicle which have less coupling with the steering and driving system and very less slippage between point of contact and the floor. The update equations for this model are as

(8) In a practical environment the pulse count received from certain encoders may indicate an over count due to workspace and operating conditions. So the least value of and estimated from the equations , (3) to (8) can be used for computing the pose of the vehicle.

6. Conclusions The DISLiB described in this paper, developed for mobile robot localization is a high resolution system which is simple, fast and accurate without much of computational burden or significant processing. Most of the localization research works are experimented in laboratory or room like environment. But most of the service mobile robot vehicles are employed in industries, warehouses etc. where a particular path is defined for their movement. Most of the available beacon's performance in corridors and narrow passages are not satisfactory but the performance of DISLiB is very encouraging in these situations. The installation is not limited to corridors, provided the vehicle crosses the light sheet for position update and error correction. Even in indoor applications the inclined paths cause localization errors which are very difficult to eliminate. But by installing DISLiBs at appropriate locations one can easily reduce the same. Normally the DISLiBs are fixed in such a way that the vehicle crosses the beacon with an angle equals to zero (i.e. = ). The bit length used for position encoding can be easily increased for mapping large workspaces. Separate firmware for encoding different beacons can be eliminated by incorporating either configuration switch inputs to the system or a RS 485 type of network and a Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

host computer. The wireless closed loop monitoring increases the overall efficiency of the system. While crossing the very short distances of around 0.12 m (ELST ), the system assumes uniform vehicle speed and accurate odometry. In order to guarantee this, the beacon installation points can be selected accordingly. For exterior mobile robot localization, the beacon systems can be organised on pillars in a side looking arrangement. The effective light sheet thickness depends on light intensity, mounting height and receiver sensitivity. The non-uniform distribution of the beacon light intensity results in slight variations of light sheet thickness on the sides of the passage. This can be reduced by the use of multiple infrared LEDs. This system will obviate the inherent odometric and INS position errors.

AUTHORS James Kurian* and P. R. Saseendran Pillai - Department of Electronics, Cochin University of Science and Technology, Cochin-682022, INDIA. Phone: +91-4842576418; Fax: +91-484-2575800. E-mails: james@cusat.ac.in, prspillai@cusat.ac.in. * Corresponding author

References [1]

[2]

[3]

[4]

[5]

[6]

[7]

[8]

[9]

[10]

Siegel M., Gunatilake P., “Remote Inspection Technologies for Aircraft Skin Inspection”, IEEE Workshop on Emergent Technologies and Virtual Systems for Instrumentation and Measurement, Ontario, Canada, May 1997. Prassler E., Ritter A., Schaeffer C., Fiorini P., “A Short History of Cleaning Robots”, Autonomous Robots, Special Issue on Cleaning and Housekeeping Robots, vol. 9, issue 3, December 2000. Iborra A., Pastor J., Alvarez B., Fernandez C., Merono J., “Robots in Radioactive Environments”, IEEE Robotics and Automation Magazine, December 2003. Stentz A., Bares J., Singh S., Rowe P., “A Robotic Excavator for Autonomous Truck Loading”, Autonomous Robots, vol.7, 1999, pp. 175-186. Rossetti M. D., Kumar A. and Felder R., "Mobile Robot Simulation of Clinical Laboratory Deliveries". In: The Proceedings of the Winter Simulation Conference. 1998. Montemerlo M., Pineau J., Roy N., Thrun S., Verma V., “Experiences with a Mobile Robotic Guide for the Elderly”. In: Proceedings of the AAAI National Conference on Artificial Intelligence, Edmonton, Canada, 2002. Bahl. R., “Object Classification using Compact Sector Scanning Sonars in Turbid Waters”. In: Proc. 2nd IARP Mobile Robots for Subsea Environments, Monterey, CA, vol. 23, 1990, pp. 303-327. Anthony J. Healey, “Application of Formation Control for Multi-Vehicle Robotic Minesweeping”. In: Proceedings of the IEEE CDC Conference, 2001. Everett H.R., Gage D.W., "From Laboratory to Warehouse: Security Robots Meet the Real World", International Journal of Robotics Research, Special Issue on Field and Service Robotics, vol. 18, no. 7, July 1999. Pastore T.H., Everett H. R., and Bonner K., "Mobile Robots for Outdoor Security Applications". In: American

Articles

[11]

[12]

[13]

[14]

[15]

[16]

[17]

[18]

[19]

[20]

[21]

[22]

[23]

[24]

[25]

[26] [27]

N° 2

2009

Nuclear Society 8th International Topical Meeting on Robotics and Remote Systems (ANS'99), Pittsburgh, PA, April 1999. Borenstein J., Everett H., Feng L., Wehe D., “Mobile Robot Positioning - Sensors and Techniques”, Journal of Robotic Systems, Special Issue on Mobile Robots, vol. 14, no. 4, 1997. Kleeman L., “Optimal Estimation of position and Heading for Mobile Robots Using Ultrasonic Beacons and Deadreckoning”. In: Proceedings of the IEEE International Conference on Robotics and Automation, Nice, France, 1992. Leonard J. J. and Durrant-Whyte H. F., “Mobile Robot Localization by Tracking Geometric Beacons”, IEEE Transactions on Robotics and Automation, vol. 7, no. 3, 1991, pp. 376-382. Wijk O., Christensen H.I., “Localization and navigation of a mobile robot using natural point landmarks extracted from sonar data”, Robotics and Autonomous Systems, vol. 31, 2000, pp. 31-42. López-Sánchez M., Esteva F., de Màntaras R.L., Sierra C., Amat J., “Map Generation by Cooperative Low-Cost Robots in Structured Unknown Environments”, Autonomous Robots, vol. 5, no. 1, 1998, pp. 53-61. Sukkarieh S., Nebot E.M., Durrant-Whyte H.F., “A High Integrity IMU/GPS Navigation Loop for Autonomous Land Vehicle Applications”, IEEE Transactions on Robotics and Automation, vol. 15, no. 3, 1999, pp. 572-578. Borenstein J., Feng L., “Measurement and Correction of Systematic Odometry Errors in Mobile Robots”, IEEE Transactions on Robotics and Automation, vol. 12, no. 6, 1996, pp. 869-880. Barshan B., Durrant-Whyte H.F., “Inertial Navigation Systems for Mobile Robots”, IEEE Transactions on Robotics and Automation, vol. 11(3), 1995, pp. 328-342. Lee S., Song J.-B., “Robust Mobile Robot Localization using Optical Flow Sensors and Encoders”. In: Proc. of the IEEE Int. Conf. on Robotics & Automation, April 2004, pp. 1039-1044. Leonard J.J., Durrant-Whyte H.F., Cox I.J., “Dynamic map building for an autonomous mobile robot”, Int. J. Robotic Research, vol. 11, no. 4, 1992, pp. 286-298. Winton, P. and E. Hammerle , “High resolution position estimation using partial pulses”, IEE Electronics Letters, vol. 36, no. 10, 2000, pp. 897-898. Lingemann K., Nuchter A., Hertzberg J., Surmann H., “High-speed laser localization for mobile robots”, Robotics and Autonomous Systems, vol. 51, 2005, pp. 275-296. Hoppenot P., Colle E, Barat C., “Off line localization of a mobile robot using ultrasonic measures”, Robotica, vol. 18, 2000, pp. 315-323. Han Y., Hahn H., “Localization and classification of target surfaces using two pairs of ultrasonic sensors”, Robotics and Autonomous Systems, vol. 33, 2000, pp. 31-41. F. Thomas and L. Ros, “Revisiting trilateration for Robot Localization”, IEEE Transactions on Robotics, vol. 21, no. 1, 2005, pp. 93-101. Fu K.S., Gonzalez R.C., Lee C.S.G, Robotics - Control, Sensing, Vision and Intelligence, McGraw-Hill, 1987. Cox I.J., “Balache-An Experiment in Guidance and Navigation of an Autonomous Robot vehicle”, IEEE Transactions on Robotics and Automation, vol. 7, no. 2, 1991. pp. 193-204.

Journal of Automation, Mobile Robotics & Intelligent Systems

[28] [29]

[30] [31]

VOLUME 3,

N° 2

2009

http://www.evolution.com/products/ Data sheet, “CYWUSB6935 Wireless USB LR 2.4-GHz DSSS Radio SoC”, Cypress Semiconductor Corporation, 2005, www.cypress.com. Peterson R., Ziemer R., Borth D., Introduction to Spread Spectrum Communication, Prentice Hall, 1995. De Cecco M., “Sensor fusion of inertial-odometric navigation as a function of the actual manoeuvres of autonomous guided vehicles”, IOP Measurement Science and Technology, vol. 14, 2003, pp. 643-653.

Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 2

2009

PROPER SELECTION OF THERMAL INSULATION MATERIALS Received 25th April 2008; accepted 12th September 2008.

Taher M. El-Shiekh, A.A. Elsayed

Abstract:

1. Major insulation materials

Thermal insulation plays a key role in the overall energy management picture. It is interesting to consider that by using the insulation, the entire energy requirements of a system are reduced. Since, the main function of insulation is to reduce the heat transfer, the insulation material must have the appropriate characteristic to retard the transport of heat occurred by conduction, convection and radiation. The proper selection and application of various insulation materials are very important in the domain of heat transfer. This paper discusses how to select thermal insulation material and on what basis a decision is made for application, concerning material selection and costs, through brief discussion about major insulation materials, properties, steps for economic selection, applications, economics of insulation and thickness, location, case study and conclusion.

Table1 shows the major insulations materials used in commercial and industrial installations and its description (1, 2, 3, 4), such as calcium silicate, glass (Fiber glass and cellular glass), mineral fiber / rock wool, expanded perlite, elastomeric foam, plastic foams, refractory and insulating cement.

Keywords: thermal insulation, insulation material, economic selection.

2. Selection process of insulation material The most appropriate insulation material is to be selected based on, application requirements, properties of the insulation material, economic thickness of insulation, design life and economics of insulation selection. 2.1. Application requirements There are two items that must always be considered to determine which insulations are suitable for service. These are range of operating temperature and location or ambient environment. The operating temperature range is classified into the following categories as shows in Table 2.

Table 1. Major insulations materials and its description.

Type of insulation material

Description of insulation material

Calcium silicate

The products, based on this, are formed from a mixture of lime and silica, reinforced with organic and inorganic fibers and moulded into rigid forms. It is suitable for temperatures from 250°F to 1000°F.

Glass (Fiber glass and cellular glass)

Fiber glass, service temperature ranges from -14.4°F up to 1200°F, 1000°F, and 850°F for glass fiber blankets, glass fiber boards and glass fiber piping covering. While cellular glass is used in temperature range -450°F to 900°F.

Mineral fiber / rock wool

The products are distinguished from glass fiber in that the fiber is formed from molten rock or slag rather than silica. Upper temperature limit can reach 1900°F.

Expanded perlite

Perlite is noncombustible and operates in the intermediate and high temperature ranges.

Elastomeric foam

Elastomeric insulations possess good cutting characteristics. The upper temperature limit is 250°F.

Plastic foams

Available in pre-formed shapes and boards, service temperature range -297°F to 300°F.

Refractory

Insulating refractory consist primarily of two types, fiber and brick. The products are used in high temperature applications.

Insulating cement

Cements may be applied to high temperature surfaces.

Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

2.2. Properties of insulation material There are many different types of insulation materials for piping applications. The insulation type's properties are temperatures range, thermal conductivity, compression strength, fire hazard classification and cell structure as in Table 3. 2.3. Economic thickness of insulation Fig. 1 shows the economic thickness of insulation (5, 6), where the total cost is quite high when insufficient insulation thickness is used. The cost drops to a minimum when the optimum thickness is used, then rises again when an uneconomical increased thickness is chosen.

N° 2

2009

Fig. 2 shows the economic steps for selecting an insulation material

3. Case study This is a case study for energy saving through cooperation between Egyptian Petroleum Research Institute and one of the Egyptian refineries located in Cairo. The Boiler Feed Water Tank (BFWT) in which hot condensate is being collected along with makes up water is fully un-insulated. The condensate line is also not insulated. Heat loss by radiation and convection is considered while conduction loss is negligible.

Table 2. Shows operating temperature range is classified into the following categories. Temperature ranges

Field of application

Suitable type of insulator

Cryogenic Cryogenic service conditions are very critical (from -455 to -150 °F) and require a well designed insulation system.

Closed-cell foamed glass (cellular glass), fiber glass, and some plastic foam.

Low temp. (from -150 to 212 °F)

Refrigeration systems used in some industries.

Glass fiber, plastic foams, phenolic foams, elastomeric materials, and cellular glass.

Intermediate temp. (from 212 to 1000 °F)

Refineries, power plants and chemical plants.

Calcium silicate, glass fiber, mineral wool, and expanded perlite.

Superheated steam, boiler exhaust ducting, High temp. (from 1000 to 1600 °F) and some process operations.

Calcium silicate, mineral wool, and expanded perlite.

Furnaces and kilns in steel mill, heat treating Refractory (from 1600 to 3600 °F) and forging shops.

Ceramic fibers are used, with alumina-silica fibers firebrick.

Table 3. Shows properties of insulation types. Insulation type and form

Temp. range (°F)

Thermal conductivity Thermal conductivity Cell structure Btu-in/hr.ft².°F (Permeability and Btu-in/hr.ft².oF at Tmean (200 °F) moisture absorption) at Tmean (75 oF)

Calcium silicate blocks, shapes and P/C Glass fiber blankets Glass fiber boards Glass fiber pipe covering Mineral fiber blocks and P/C Cellular glass blocks and P/C Expanded perlite blocks, shapes, P/C Urethane foam blocks and P/C Isocyanurate foam blocks and P/C Phenolic foam P/C Elastomeric closed cell sheets and P/C MIN-K blocks and blankets Ceramic fiber blankets

to 1500

0.37

0.41

Open cell

to 1200 to 1000 to 850

0.24 - 0.31 0.22 0.23

0.32 - 0.49 0.28 0.3

Open cell Open cell Open cell

to 1900

0.23- 0.34

0.28-0.39

Open cell

-450 to 900

0.38

0.45

Closed cell

to 1500

0.46

Open cell

(-100 to -450 ) to 225 to 350

0.16 0.18

95% Closed cell

0.15

93% Closed cell

-40 to 250 -40 to 250

0.23 0.25 0.27

_ _

Open cell Closed cell

to 1800

0.19 0.21

0.20-0.23

Open cell

to 2600

Open cell

P/C means pipe covering. Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 2

2009

3.2. Heat loss calculation for condensate line In this section we calculate the annual equivalent energy loss through un-insulated pipes, hence the annual cost. The length and surface area of the condensate pipe is given in the Table 4, these data collected from one of the Egyptian refineries located in Cairo.

Step 1:

Measure the area of the surface to be insulated and its temperature

Step 2:

Calculate the quantity of fuel and equivalent cost lost due to non-insulation

Step 3:

Specify temperature range, required life time and any limiting factors

Step 4:

Identify the suitable insulating material that meets all the requirements

Step 5:

Find out the correct size of Insulation thickness

Step 6:

Work out the initial cost, installation cost, maintenance costs, interest on initial cost

Step 7:

Work out the economic comparisons of attractive systems

Fig. 1. Economic thickness of insulation. 3.1. Heat loss calculation for BFWT The Boiler Feed Water Tank (BFWT) is fully uninsulated and with the following conditions. Average surface temperature of the bfwt, T1 = 55°C = 131°F = 590.6°R. Ambient Temperature, T2 = 30°C =86°F = 545.6°R. No. of sides exposed to the atmosphere = 5 (i.e., except bottom), (Length = 7.546 ft; Width = 7.546 ft &Height=6.561 ft), A = (2 × 7.546 × 6.561)) + (2 × 7.874 × 6.561) + + (7.546 × 7.874) = 261.755 ft². Total heat loss per ft² area can be calculated from the following equation (1):

if step 2 cost > step 6 cost

If No

If Yes Select most economically attractive system

Where = rate of heat loss, = surface area of the tank, (A=261,755 ft²), = average temperature of pipe surface, 55 °C = 131°F = 590.6 °R, = average temperature of the surroundings (air), 30 °C = 86°F = 545.6 °R, = surface emittance, 0.9.

Step 8:

Hence, Q/A = 86.259 Btu/ hr.ft². Total heat loss, Q = 86.259 ×261.755 = 22578.74 Btu/hr. The BFWT is work as 350 day/year (15 days to maintenance). Then the annual equivalent energy loss through BFWT = 22578.74 × 24 × 350 = 189661416 Btu. Changing the annual equivalent energy loss through BFWT to equivalent furnace oil as following; Equivalent furnace oil = (189661416×0.252) / (9650 x 0.75) = 6603.75 kgs/year = (6603.75 / 0.92) = 7177.99 Liters/year. Annual cost equivalent at 3.4 L.E. (Egyptian Pound) per liter = 7177.99 × 3.4 = 24405.18 L.E.

Table 4. Shows diameters, lengths and surface areas of the condensate pipe.

Articles

Step 9:

Do the insulation work

Fig. 2. Economic steps for selecting an insulation material.

Size of D, inches Length L, ft 2 4.59 1.5 56.43 1.5 267.057 1 222.11 Total Area A1

Area (pDL) 2.403 22.16 104.873 58.148 187.584

= total surface area of the pipes, (A=187.584ft2), = surface emittance, 0.9,

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

= average temperature of pipe surface, 60 °C = 140°F = 599.6 °R, = average temperature of the surroundings (air), 30 °C = 86°F = 545.6 °R. Heat loss through un-insulated pipe = Convection Loss + Radiation Loss. Hence, Q/A = 79.723 Btu/ hr.ft2. Total heat loss, Q = 79.723 ×187.584 = 14954.759 Btu/hr. Annual equivalent energy loss through pipes = 14954.759 × 24 × 350 = 125619977.5 Btu. Equivalent Furnace Oil = (125619977.5 × 0.252) / (9650 x 0.75) = 4373.918 kgs/year = (4373.918/0.92) = 4754.259 Liters/year. Annual cost equivalent at 3.4 L.E. (Egyptian Pound) per liter = 4754.259 × 3.4 = 16164.481 L.E. Total cost = 24405.18 + 16164.481 = 40569.66 L.E.

N° 2

2009

Respecting to above factors, we may select fiber glass as insulation material, which can be used in the temperature range of -14.4°F to 850°F. From Table 5 we can select the thickness of insulation material as given below. In this application, the temperature range is more than (+ 16.2°C) 61, 16°F, we can go for 1 inch (25 mm) as insulation thickness. Now, we have to work out the initial cost, installation cost, maintenance cost and interest on initial cost: Initial and installation cost for Fixing the glass fiber mattress having density of (100 3 Kg/m ) 25 mm thick with wire netting and Aluminum jacket. Sealing all joints with sealing compound if the pipelines are exposed outside is given in the following Table 6. Using simple payback method to evaluate insulation investment, Simple payback period = (10619.03 /40569.66) x 12 = 3.14 Months. Hence, from the above total insulation cost and annual heat loss cost, we are getting very attractive saving cost. So, we can select glass fiber as the insulation material.

3.3. Selecting the insulation material and calculationof total cost Properties of insulation materials should be high compressive strength, resistant to moisture absorption and chemical, non combustible and not expensive. Temperature range = 30°C to 75°C and required life time = 5 years.

Table 5. Shows thickness of insulation material based on temperature range and diameter. Temperature ranges in °F

Insulation thickness

Pipe diameter in inches

inches

Below (- 30.2)

0.59 to 3.1496

3.397

100

From (-30.02) to (+39.02)

3.937 to 11.811

4.921

125

0.59 to 3.1496

2.953

From (+39.92) to (+55.4)

All sizes

1.968

From (+56.12) to (+60.08)

All sizes

1.575

From (+61.16) and above

All sizes

0.984

Table 6. Shows the total of insulation cost. Items

Unit

(Price / unit) L.E.

Total price L.E.

Boiler feed water tank

261.755 ft²

4188.08

Condensate pipe 2 inches

4.59

8.5

39.05

Condensate pipe 1.5 inches

323.48

8.3

2684.88

Condensate pipe 1 inch

222.11

7.75

1721.35

Total initial and installation cost

8633.36

Interest at 12%

1036

Annual maintenance at 11%

949.67

Total cost

10619.03 Articles

Journal of Automation, Mobile Robotics & Intelligent Systems

4. Conclusions The appropriate insulation materials must be selected based on temperature, thermal conductivity and other limiting factors that might limit application. The appropriate thickness must be determined for the particular application. While doing the payback calculation on insulation, one has to consider the capital cost, interest on capital cost, depreciation period and maintenance cost. Comparing the total insulation cost and annual heat loss cost, we are getting very attractive saving cost for the case study. So, we can go for 1 inch (25 mm) as insulation thickness from glass fiber mattress having density of 3 (100 Kg/m ). The simple payback period to evaluate insulation investment is 3.14 Months.

AUTHORS Taher M. El-Shiekh* - Egyptian Petroleum Research Institute, Process Design and Dev. Dept. Nasr City, Cairo, Egypt. Email: taherelshiekh@yahoo.com. A. A. Elsayed - AL Azhar University, Faculty of Eng., Petroleum Dept. Nasr City, Cairo, Egypt. Email: sayed_669@yahoo.com. * Corresponding author

References [1] [2] [3] [4]

[5] [6]

Turner W.C., Malloy J.F., Thermal Insulation Handbook, McGraw-Hill: New York, 1981. Waldo J., Piping Handbook, Chapter B7, Sixth Edition, McGraw-Hill: New York, 1992. Harrison M.R., Energy Management Handbook, Chapter 15, John Wiley & Sons: Canada, 1992. Kanakia M., Herrera W., and Hutto F., “Fire Resistance Tests for Thermal Insulation”, Journal of Thermal Insulation. Apr. , 1978, Technomic, Westport. Conn. McMillan, “Fuels Steam Power”, Trans. Am. Soc. Mech. Engrs., vol. 51, 1929, pp. 349-355. McAdams W. H., Heat Transmission, McGraw-Hill: New York, 1954.

Articles

VOLUME 3,

N° 2

2009

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

NÂ° 2

2009

A NEW DESIGN OF A ROBOT PROTOTYPE FOR INTELLIGENT NAVIGATION AND PARALLEL PARKING Received 6th November 2008; accepted 6th March 2009.

Chokri Abdelmoula, Fakher Chaari, Mohamed Masmoudi

Abstract: Nowadays, the design of industrial vehicles and movable cars is based on the automation of their different tasks, which are currently handled by humans. These tasks, such as maneuvering robots in complex environments, require high level of precision that cannot be guaranteed by humans. Manual operations are likely to produce errors of computation and optimization of navigation and manoeuvre (left, right, veeringâ&#x20AC;Ś). In this paper, a novel prototype of a well-structured robot for intelligent navigation and parallel parking applications is presented. The robot have two axels, the front one is composed of two wheels that are manoeuvred by a stepper motor, and a pinion rack system for controlling the rotation of the wheels, and also the orientation of the robot. The driving wheels are mounted in the rear axle of the robot and are commanded by two DC motors. The design allows modification of the robot structural components whenever required. In addition to the mechanical components, the prototype is equipped with a DC power supply, three infra-red sensors, one ultrasound sensor, and control modules composed of an FPGA card, microcontroller card and two cards which are responsible for commanding actuators. The parameters of the mechanical and electronics components are optimised to perform multiple tasks for training and instruction applications. A mathematical model that describes the dynamics of the robot prototype is also developed. Simulation, experimental and theoretical investigations were carried out consisting in navigation and parallel parking manoeuvres. It was confirmed that the experimental and theoretical results agree well in both applications. Keywords: autonomous vehicles, embedded design, and mobile robot.

1. Introduction Creating autonomous robots is one of the main up-todate research activities. Real-world mobile robots operate in an environment that is not engineered for a particular robot. In other words, prior knowledge of the environment is limited and unreliable because of the complexity and unpredictable dynamics of the surroundings. Consequently, the ability of a robot to plan its motion autonomously is of vital importance. In the literature, several research studies addressed the problems of navigation and intelligent car parking, obstacle avoidance and target tracking control of mobile robots. The control of a mobile robot in dynamic and unstructured environments typically requires efficient processing of data/ information to ensure precise navigation and intelligent parallel parking.

1.1. Motivation The present study aims at developing a novel approach of obstacle avoidance for a mobile robot that uses a hierarchical control for autonomous navigation in an unknown environment. A new configuration of hybrid architecture for intelligent parallel parking with the use of a PC for command is also developed. For control purposes, an RF module installed on the chassis of the robot is used. A mathematical model, which describes the robot kinematics, is incorporated into the proposed design to control the robot within its unknown environment. The collected sensors information regarding this environment is processed to guide the mobile robot to its target while avoiding collisions. For example, when more than two parking lots with different widths are available, the robot measures the width of each lot, and thus, selects the suitable place for parking. Using the information's of the sensors, the robot parks while avoiding obstacles in all directions (forward, backward, right and left). To confirm its efficiency, the proposed control approach is implemented on a robot prototype. Software for the microcontroller is developed in order to acquire the sensor data and perform algorithms and functions that control the speed of the two rear DC motors and the direction of the mobile robot. Both scale and weight of the platform are optimised during the manufacturing stage of the robot prototype. The prototype addresses the safety issues and incorporates the interfacing modules and troubleshooting of the software with the hardware. Algorithms for autonomous navigation or intelligent parking are implemented. This paper is structured as follows. Section II presents the description of the mobile robot in its new design. Section III discusses the proposed architecture of the robot prototype. In section IV the mathematical model describing the kinetic behaviour of the mobile robot is presented. Section V presents the proposed control approach. Section VI provides simulations and real-time implementation results as well as their comparison with the experimental results. The last section of the paper concludes the work with suggestions for future studies. 1.2. Background/Literature review Recently, there has been a growing interest in parallel-parking problems of car-like mobile robots for navigation [5, 9-15]. Lyon [13] formulated the parallel-parking problem of curvilinear path generation for a car with nonholonomic constraints, where the slope and curvature constraints are used to derive a fifth-order polynomial equation denoted as the parallel-parking path. The relationship between the rear-path curvature and the Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

steering angle was also developed. Reactive type-2 fuzzy architecture, for the real-time control of mobile robots navigation in dynamic unstructured indoor and outdoor environments, is one of many research activities in mobile robots that use interval type-2 Fuzzy Logic Controller (FLC) to implement the basic navigation behaviours and their coordination to produce a type-2 hierarchical FLC [1]. Simon [2] developed an embedded fuzzy controller for a nonholonomic mobile robot that was built based on the behaviour-based artificial intelligence, where several levels of competences and behaviours were implemented. A hybrid learning approach for a neuro-fuzzy system applied to obstacle avoidance of a mobile robot was suggested by Meng and Deng [3] to prove that the simulation environment is supervised via learning. Using the hybrid learning approach, an efficient and compact neuro-fuzzy system can be generated for obstacle avoidance of a mobile robot in the real world. Many other studies investigated the computing-based-embedded design of an intelligent wall/lane following vehicle. For unstructured environments, soft computing techniques for vehicle control systems were modelled using highly non-linear differential equations. Using these techniques, two intelligent controllers based on fuzzy logic and artificial neural network were designed for performing a wall following task [4]. These controllers were contrasted for hardware resource requirements, operational speed, and trajectory tracking errors. Many others problems were encountered in testing forward navigation and the direction control module of the platform [6]. Typical autonomous vehicles employ conventional controllers that generally require the use of complex mathematical models [16-17]. However, these controllers are prone to many issues such us modelling uncertainties in real-world environment and processing of noisy sensor information. Yang and Meng [18] presented a real-time collision-free motion-planning algorithm using a Neuronal-Network (NN) based approach. Unlike conventional NN models, dynamics of each neuron is characterized by a shunting or an additive equation derived from Hodgkin and Huxley's [19] membrane equation. Weights are predefined at the stage of NN design, eliminating the need for a learning procedure. The proposed algorithm demonstrated effectiveness in parallel parking and navigation applications. Zalama et al. [20] designed an NN model for navigation. Reinforcement learning is employed in which the control system learns the appropriate action through a credit and punishment system that assigns rewards and penalties according to the correctness of the control action. The goal is to maximize the long-term reward. Based on this technique, the control system successfully learnt collision avoidance, wall following, and goal-reaching behaviours. In terms of computing medium, many researches utilize microprocessors and FPGA (Field Programmable Gate Array) based hardware. The design and implementation of a real-time fuzzy logic based parallel parking system was proposed with two control units consisting of a main controller and a secondary fuzzy logic controller [5]. The latter is employed for realizing the wall following 48

Articles

N째 2

2009

task, which plays a key role in the parking system. Based on performance and flexibility considerations, the control units were implemented onto a reconfigurable hardware platform. Krohling et al. [21] used an evolutionary algorithm, implemented on FPGA, to provide navigation for an autonomous robot in unknown and changing environments. Fakhfakh et al. presented a work that show different analysis (Numerical and Experimental) of a Gear System with Teeth Defects [22]. Few papers were interested on the kinetic behaviour of the mobile robot influenced by the driving system and the environments constraints. Also the optimization of the manoeuvring operation such as changing direction is not well tested in literature. In the other hand, existing mobile robots have specific operations objectives, which are induced by their limited design. The present work intends to design and implement a new generation of robots able to solve manoeuvring operations in complex situations using an original driving system and orientation mechanism.

2. Description of the mobile robot

Fig. 1. Autonomous mobile robot. A structure of a mobile robot (Fig. 1) is developed for navigation control and intelligent parallel parking in a dynamic environment. The robot is of length 350 mm, width 240 mm and weight 2,840 kg. The hardware and software of the proposed robot prototype were developed in the Micro Electro Thermal Systems Laboratory of the National School of Engineers in Sax, Tunisia. The prototype was properly designed to address the concerns of automobile manufacturers raising problems of intelligent parallel parking and navigation that require scheduled tasks with/without human intervention in a complex environment. The robot is trained to use artificial vision and is programmed to follow trajectories in its environment by acquiring information from the positioning sensors. It is also talented to move autonomously in its unknown environment, to avoid obstacles, and to perform intelligent parallel parking using four sensors mounted on the chassis of the robot. As it detects obstacles in its environment, the robot measures its position with respect to the parking place. This prototype, illustrates many types of smart applications, such as universal and didactic applications [23]. The developed mobile robot consists of the following

Journal of Automation, Mobile Robotics & Intelligent Systems

modules (Fig. 2): Four positioning sensors S1-S4 responsible for detecting obstacles, An electronic card (Card1) for controlling the stepper motor (M3), and the DC motors (M1 and M2), A micro controller (Card2) for database acquisition, An F.P.G.A card (Card3) for processing the data supplied by the three infrared sensors and the ultrasonic one, Front and rear axles, Four wheels w1-w4, battery, one rack in front axle for manoeuvring, and two reducers R1 and R2.

Fig. 2. Schematic of the mobile robot. Figures 3 and 4 show the components of the front and rear axles, respectively. The front axle consists mainly of two wheels, a pinion rack system for manoeuvring the robot during navigation, and a stepper motor. To change the direction, the DC motors operate at different angular velocities. The rear axle consists mainly of two gearboxes independently mounted on the chassis driven by two DC motors. This driving system is essentially used for the system motion and also for the direction changing manoeuvres.

Fig. 3. Front axle of the mobile robot.

VOLUME 3,

N° 2

2009

3. Proposed Architecture of the Robot Prototype The architecture design of the proposed prototype is valid for various fields of application that use FPGA cards for intelligent parking in different working conditions. In fact two mains design levels for the robot prototype are encountered in literature [7]: the global and physical levels. The global level is defined by a set of exteroceptif sensors and the physical level by the mechanical system where the proprioceptif sensors are integrated. For planning tasks, the exteroceptif sensors: compute the shortest track between two memorized locations, accomplish short cuts to unexplored zones within the working environment, memorize the robot trajectory, and foresee a trajectory between the two locations. The robot prototype integrates hybrid controllers for building intelligent algorithms using an FPGA card for navigation and parallel parking applications. [7-8] The different tasks of planning and obstacle detection by the proprioceptif sensors in the proposed platform are converted into data that are processed by the FPGA card to activate the actuators without disruption during other decisions. The development system allows the integration of multi-tasks operation for navigation in unknown environments Two basic application concept of the proposed robot in an unfamiliar environment containing fixed or movable barriers were developed. In the first one, the robot navigates in this environment while detecting the distances that separate the robot platform from barriers. The decision to change direction (right or left) depends on the distance that separates the platform of the detected obstacle. In the second application, several parking stations are available and the robot have to select the suitable one, which fits its width, to travel through it.

4. Kinematics model of the robot In order to describe and realize its motion, we develop expressions for the kinematics of the robot prototype. A stepper motor geared to a rack commands the front axle, represented in Fig. 5. As the motor rotates, the rack moves to the right or to the left depending on the direction of rotation. On each side of the front axle, a wheel mechanism composed of three joints ensures the rotation

Fig. 4. Rear axle of the mobile robot. Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

2009

of the wheels (Fig. 6). The configuration of the mobile robot during navigation is shown in Fig. 7. We distinguish three position vectors R , Rr and Rl that locate the centres of the front axle, right and left wheels, respectively. L is the distance that separates the centres of front and rear wheels, E is the axle length, R is the radius of gyration, and a is the

steering angle such that sin a = L/R. For a given steering angle, the rack displacement is characterized by the arc length AA'. The norm of the position vector locating the centre of joint A can be expressed by

(1)

Fig. 5. Schematic of the front axle.

Fig. 6. Wheel mechanism (a) non-rotated (b) rotated through and angle.

Fig. 7. Robot configuration during curving motion. 50

N째 2

Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 2

2009

Thus, the arc length travelled by this joint is as follows:

(2)

J3 2 3

Where d represents the translation distance of the rack, hence, the angular rotation of the pinion and then of the stepper motor is given by

J4 4

(3) Where dp is the pinion pitch diameter, m is the module, and Zp is the pinion tooth number. In case of a left turn, the left wheels of the front and rear axles travel a shorter distance as opposed to the right ones. This can be justified by the fact that Rr is less than Rl. For example a rotation of 90° corresponds to the following distances travelled by the right rear and front wheels respectively: (4) In the absence of a differential gears system at the rear axle, the DC motors are operated independently to regulate the angular velocities of the left and right wheels. The time to travel a gyration angle of 90° can be expressed by:

J6 6

Fig. 8. Kinematics of the rear axle. The front wheels and the gearboxes mounted on the rear axle constitute the principal mechanism for navigating the mobile structure in its working environment. As the robot detects an obstacle, it makes a turn, as depicted by Fig. 9. We allow a moving frame (O, X, Y) to be attached to the centre point of the front axle. B( )

(5) Where Vl and Vr are respectively the linear velocities of the left and right wheels, and V is the linear velocity at the robot centre line. Hence, Vl and Vr can be expressed by and

arc

(6)

Consequently, in order to ensure a regular gyration of the robot, the angular speeds of the left and right wheels have to be reduced by and increased by

respectively. Robot navigation for intelligent parking with a gyration angle of 90° requires the computation of the steering angle, the distance travelled by the rack, the angular rotation of the stepper motor, and the speeds of the DC motors. This enables precise parking by commanding the two DC motors with independent speeds. A model for the kinematics associated with the rear axle is represented in Fig. 8, where Cm and Cw denote the torques associated with the motor and wheel, respectively. The robot torque is regulated by a control law, which is automatically constructed by intelligent algorithm. - is the torsional stiffness of the different shafts and the gear mesh between the different gears. kg(t) - is modelled by a time varying mesh stiffness in order to take into account the mesh process. Ji (i=1…..6) are the inertia moments of the gears. i (i=1…..6) are the different angular rotation of the gears.

arc(t)

X(t)

Y(t)

A( 0)

Fig. 9. Trajectory of the robot for obstacle avoidance. During an angle turn (t) with respect to the Y axis, the coordinates of the centre point of the front axle are

(7)

(8)

where rw is the wheel radius, g is the angular velocity of the robot around o1, and (t) is the angular position of the rotating wheel. In all cases, the computed trajectories must start and finish at zero-curvature configuration.

5. Proposed control Approach We propose the use of classical and hybrid control strategies for intelligent navigation and parking, respectively. Fig. 10 displays the global and physical levels of Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

In order to show the efficiency of the proposed approach, the intelligent navigation and parking applications are simulated in this section.

Pilot system

Processed Information

Collection of information Perception (Sensors)

Decisions

Orders Action (Motors)

Fig. 10. Proposed control approach. Starting from the perception by the sensors at the global level, the data is processed in a sequential way to operate the actuators at the physical level. The perception phase collects the information and processes it by a pilot system, as shown in Fig. 10. This process is characterized by sequential decisions for the control of actuators that obey the orders and instructions of the algorithm. Consequently, this approach follows the bottomup/top-down design approach. The proposed approach is characterized by control tools of a hierarchical nature, where we use the classical approach to validate the robot prototype. After the design of the mechanical components, we simulated and tested the navigation application and simulated the parking application. These applications are commanded at distance by a PC since the prototype is equipped with an RF module. Therefore, the proposed approach constitutes a phase of real actions are simulated.

6.1. Navigation Fig. 11 configures a navigation application in which the prototype has to avoid a set of obstacles within its working environment. The navigation control approach consists of an algorithm for detecting an obstacle and veering to avoid a collision without making any stop. When the robot detects an obstacle, it measures the distances on its right and left, and veers at an angle of 90째 to the left or to the right, depending on the longest distance from the obstacle. In classical robot navigation, the robot stops in front of the obstacle, measures the distances on the right and left, and then resumes its motion. Let us show a real example of this experience. An example of this navigation process is presented in Fig. 12, where Snapshots of the prototype location and orientation are associated with phases 1, 2 and 3. Figures 13a-c present simulations of this navigation case. The robot prototype starts moving along the X-axis and then makes a right turn as it detects an obstacle, which is either predictable or unpredictable. In the second phase, we notice that the translation in the X-Y plane is accompanied with a vibratory motion. This latter is mainly due to the flexibility of the gears mounted on the rear axle. As the robot navigates along either the X-axis or Y-axis, these vibrations become negligible, and due to the torsional kinematics model it is not possible to get representation of these vibrations in the X-Y plane. However, during the veering phase the dynamics of the rear axle induce important vibrations to the rigid-body motion. Fig. 13c shows the controlled trajectory of the mobile robot in X-Y plane. These simulation results illu-

Fig. 11. Configuration of a navigation in an unknown environment. 52

Articles

2009

6. Computer simulations and real-time implementation results

each of these control strategies.

Command of the mobile robot

N째 2

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 3,

N째 2

2009

Fig. 12. Implementation experimental results of navigation.

Fig. 13a. Simulation results: Phases (1, 2).

Fig. 13b. Simulation results: Phases (2, 3). Articles

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 3,

N째 2

2009

Fig. 13c. Comportment of the mobile robot in the X and Y axis. strate the efficiency of the proposed control algorithm when the mobile robot confronts obstacles. Fig. 14 shows 3D representation of the mobile robot position as a function of time.

ciated with phases 1, 2, 3 and 4 in the two appli-cations. Figures 17a and 17b display the simulated trajectories along the X and Y axes for application 1 presented in Fig. 15a. The commanded trajectory of the prototype in the X-Y plane is shown in Fig. 17c. Once more, we observe that the flexibility of the gears mounted on the rear axle induces important vibrations. Figures 18a and 18b are simulations of the commanded displacements along the X and Y axes, respectively, associated with application 2 represented in Fig. 15b. These simulations illustrate the efficiency of the proposed control algorithm when the mobile robot confronts obstacles. The results of both applications confirm the efficacy of the proposed control algorithm for commanding intelligibly the mobile robot, equipped with an RF module, at distance using a PC.

7. Conclusions and Future Studies Fig. 14. 3D representation of the commanded navigation. 6.2. Parallel parking In order to show the feasibility of the proposed algorithm for navigation, we consider the parallel parking application. A algorithm is implemented for this purpose to handle different situations. Figures 15a and 15b configure two different parallel parking applications. In the first one, the robot, whose width is 240 mm, moves on a straight path while inspecting on its right side the sizes of two parking lots (150 and 300 mm). This inspection yields decision of the correct lot where the prototype is to be parked. The mobile robot goes back a distance of 275 mm to take a turn to the right followed by a straight line leading to the selected parking lot. An example of these two operations of parallel-parking process is presented in Fig. 16a and 16b, where Snapshots of the prototype location and orientation are asso54

Articles

In this paper a novel design of an autonomous mobile robot for navigation and parallel parking applications is manufactured and implemented. The proposed robot is equipped with an FPGA card, a microcontroller module, an electronic driver module, an ultrasonic sensor and infrared sensors. These robot present real advantages: it offers optimised manoeuvring operation allowed by the novel mechanism and hardware design. In order to show the viability of the proposed design, we considered one case of navigation and two cases of parallel parking. For each case, we implemented an intelligent algorithm for obstacle avoidance. It was found that the simulated trajectories of the commanded robot show similar performance with real navigation applications. The main outcome of this research was to develop an electromechanical system that is commanded at a distance and navigates intelligently in a working environment with predictable and unpredictable obstacles. The proposed design can be easily implemented in industrial environments, such as carrying objects. It can also be used in vehicles

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 3,

N째 2

2009

Fig. 15a. Parallel parking: application 1.

Fig. 15b. Parallel parking: application 2.

Fig. 16a. Implementation experimental results (parallel parking: application 1). Articles

Journal of Automation, Mobile Robotics & Intelligent Systems

Fig. 16b. Implementation experimental results (parallel parking: application 2).

Fig. 17a. Displacement along the X-axis (parallel parking: application 1).

Fig. 17b. Displacement along the Y-axis (parallel parking: application 1). 56

Articles

VOLUME 3,

N째 2

2009

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 3,

N째 2

2009

Fig. 17c. Robot trajectory on the X -Y plane (parallel parking: application 1).

Fig. 18a. Displacement along the X-axis (parallel parking: application 2).

Fig. 18b. Displacement along the Y-axis (parallel parking: application 2). Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

for intelligent parallel parking. In addition, the prototype will be used for educational and research purposes at undergraduate and graduate laboratories in the Department of Electrical Engineering of the National Engineering School of Sfax. Future research will emphasize the integration of a mobile camera, instead of a set of sensors, into the mobile robot for intelligent navigation and obstacle avoidance. This camera is used not only for detecting obstacles but also for positioning the robot within its working environment. The implementation of intelligent control approach based on genetic algorithms with the use of an FPGA module card will be also investigated.

[9]

[10]

[11]

[12]

AUTHORS Chokri Abdelmoula* - EMC Research Group, National Engineering School of Sfax, BP 1173 3038 Sfax - Tunisia, tel: 216-4-244423. E-mail: Chokri_abdelmoula@yahoo.fr. Fakher Chaari - DMS Research Unit, National Engineering School of Sfax, BP 1173 3038 Sfax - Tunisia, tel: 216-4274088. E-mail: Fakher.chaari@gmail.com. Mohamed Masmoudi - EMC Research Group, National Engineering School of Sfax, BP 1173 3038 Sfax - Tunisia, tel: 216-4-274088. E-mail: Mohamed.masmoudi@enis.rnu.tn. * Corresponding author

[13]

[14]

[15]

[16]

References [1]

[2]

[3]

[4]

[5]

[6]

[7]

[8]

Hagras H.A., “A hierarchical Type-2 Fuzzy Logic Control Architecture for Autonomous Mobile Robots”, IEEE Transactions on Fuzzy Systems, vol. 12, no. 4, 2004, pp. 524-539. Simon X. Y., Hao L., Max Q.-H. M., Peter X. L., “An Embedded Fuzzy Controller for a Behavior-Based Mobile Robot with Guaranteed Performance”, IEEE Transactions on Fuzzy Systems, vol. 12, no. 4, 2004, pp. 436-446. Meng J. E., Chang D., “Obstacle Avoidance of a Mobile Robot Using Hybrid Learning Approach”, IEEE Transactions on Industrial Electronics, vol. 52, no. 3, 2005, pp. 898-905. Tsui W., Masmoudi M.S., Karray F., Song I., Masmoudi M., “Soft Computing Based Embedded Design of an Intelligent Wall/Lane Following Vehicle”, IEEE/ASME Transactions on Mechatronics, vol. 13, 2008 Issue1, pp. 125-135. Masmoudi M.S., Tsui W., Song I., Karray F., Masmoudi M., Derbel N., “Implementation of a Real-Time FPGABased Intelligent Parallel Parking System”, Journal of Advanced Computational Intelligence and Intelligent Informatics, JACIII, HNICEM 2007 Special Issue, vol. 12, no. 4, 2008, pp. 348-354. Kesselman G., Ng A., Nowostawski J., “Design and Implementation of a Vehicle Trailing Control System”, Rodolfo Reyes, 2003, www.mts.net/~png/Thesis/Pro-posal.pdf. Fitzpatrick P., “Behaviour-Based Control in Mobile Robotics”, University of Limerick, Ollscoilluimnigh, Limerick, Ireland. Available at: http://people.csail.mit.edu/ paulfitz/pub/fitzpatrick96behaviour.pdf Shih J.C., Tzuu H.S.L., “Design and Implementation of Fuzzy Parallel Parking Control for a Car-Type Mobile

Articles

[17]

[18]

[19]

[20]

[21]

[22]

[23]

N° 2

2009

Robot”, Journal of Intelligent and Robotics Systems, vol. 34, no. 2, 2002, pp. 175-194. Gómez-Bravo F., Cuesta F., Ollero A., “Parallel and Diagonal Parking in Nonholonomic Autonomous Vehicles”, Engineering Applications of Artificial Intelligence, vol. 14, no. 4, 2001, pp. 419-434. Divelbiss A.W., Wen J.T., “Trajectory tracking control of a car-trailer system”, IEEE Transactions. Control System Techno, vol. 53, 1997, pp. 269-278. Leitch D., Probert P.J., “New Techniques for Genetic Development of a Class of Fuzzy Controller”, IEEE Transactions. on Systems Man Cybernet, vol. 28, 1998, pp. 112-123. Lian K.Y., Chin C.S., Chiang T.S., “Parallel Parking a CarLike Robot Using Fuzzy Gain Scheduling”. In: Proc. of 1999 IEEE Internat. Conf. on Control Applications, vol. 2, 1999, pp. 1686-1691. Lyon D., “Parallel Parking with Curvature and nonholonomic constraints”. In: Proc. of '92, Symposium on Intelligent Vehicles, Detroit, MI, 1992, pp. 341-346. Ohkita M., Mitita H., Miura M., Kuono H., “Travelling Experiment of an Autonomous Mobile Robot for a Flush nd Parking”. In: Proc. of the 2 IEEE Conf. on Fuzzy Systems, San Francisco, CA, USA, vol. 2, 1993, pp. 327-332. Paromtchik I.E., Laugire C., “Motion Generation and Control for Parking an Autonomous Vehicle”. In: Proc. of '96 IEEE Conf. on Robotics and Automation, Minneapolis, MN, vol. 4, 1996, pp. 3117-3122. De Santis R., “Path-Tracking for Car-like Robots with Single and Double Steering”, IEEE Trans. Veh. Technol, vol. 44, no. 2, 1995, pp. 366-377. Stotsky A., Hu X., “Control of Car-Like Robots Using Sliding Angle Observers for Steering Angle Estimation”. th In: Proc. of 36 Conf. on Decis. Control, vol. 4, 1997, pp. 3648-3653. Yang S., Meng M., “Real-Time Collision Free Motion Planning of a Mobile Robot Using a Neural Dynamics Based Approach”, IEEE Trans, Neural Netw, vol. 14, no. 6, 2003, pages 1541-1552. Hodgkin A.L., Huxley A.F., ”A quantitative description of membrane current and its application to conduction and excitation in nerve”, Amer. J. Physiol, vol. 117, 1952, pp. 500-544. Zalama E., Gomez J., Paul M., Peran J.R., “Adaptive behaviour navigation of a mobile robot”, IEEE Trans. Syst. Man Cybern. Part A: Syst. Humans, vol. 32, no. 1, 2002, pp. 160-169. Krohling R., Zhou Y., Tyrrell A., ”Evolving FPGA-based robot controllers using an evolutionary algorithm”, First Int. Conf. on Artificial Immune Systems ICARIS'02, th th University of Kent, Canterbury, UK, 9 -11 September 2002. ISBN 1902671325, 9781902671321. Fakhfakh T., Chaari F., Haddar M., “Numerical and Experimental Analysis of a Gear System with Teeth Deffects”, International Journal of Advanced Manufacturing technology, vol. 25, no. 5-6, 2005, pp. 542-550. Abdelmoula C., Masmoudi M., “Robot Mobile à Usage d'Applications Universelles et Didactiques”, Brevêt d'Invention, INNORPI, June 2008.

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

N째 2

SPECIAL ISSUE SECTION

AH/EHW - the State of the Art and the Prospectus for Future Development Editors: Mircea Gh. Negoita Sorin Hintea

2009

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 3,

N째 2

2009

Editorial Oscar Castillo*, Patricia Melin

Special issue section on AH/EHW - the State of the Art and the Prospectus for Future Development With great pleasure, we would like to welcome you to this Guest Invited Issue of the JAMRIS - Journal of Automation, Mobile Robotics & Intelligent Systems. The motivation of its selected topic - AH/EHW - the State of the Art and the Prospectus for Future Development is deeply justified. Bio-Inspired Computing Technologies led to the spectacular progress of the Computational Intelligence (CI) nowadays and to its implementations in form of the Hybrid Intelligent Systems (HIS). Evolvable Hardware (EHW) has emerged as a novel and highly diversified technology and paradigm supporting the design, analysis and deployment of the high performance intelligent systems. The intellectual landscape of EHW is enormously rich. The discipline of EHW brings together hardware implementation of the main technologies of CI including fuzzy sets, neural networks, and evolutionary optimisation. But EHW systems use more than just the three broad areas mentioned above. They also cover novel areas as Artificial Immune Systems and DNA computing. The strength of EHW hinges on the synergy between these technologies supported by the advanced analogue and digital programmable circuits. This synergy helps exploit the advantages of the contributing technologies while reducing their possible limitations. The advanced programmable circuits confer the suitable hardware environment for a CI implementation from day to day more close to the intelligence of a human being. Nevertheless, it is hardware implementation of the most benefit for the society and indeed most revolutionizing application of Evolutionary Computation (EC) by leading to the so-called Evolvable Hardware (EHW). These new EC based methodologies make possible the hardware implementation of both genetic encoding and artificial evolution, having a new brand of machines as a result. This type of machines is evolved to attain a desired behaviour that means they have a behavioural computational intelligence. There is no more difference between adaptation and design concerning these machines, these two concepts representing no longer opposite concepts. A dream of technology far years ago currently became reality: adaptation transfer from software to hardware is possible by the end. Much more, the electronics engineering as a profession was radically changed: the most based on soldering assembling manufacturing technologies are largely replaced now by programming circuitry-based technologies, including EHW technologies. EHW is a special case of the adaptive hardware, namely being strongly related to the Adaptive Systems (AS) and the Adaptive Hardware (AH). The progress in EHW is rapid. The individual technologies evolve quite quickly paving a way to new interesting and truly amazing applications. In the heart of all of those is the principle of hybridisation. EHW is suitable for the dramatic changes that happen in the relation between hardware and the application environment. This is in the case of malicious fault/defects and need for new emergent functions that claim for in-situ synthesis of a totally new hardware configuration. It is not surprising at all witnessing a lot of activities and achievements within this realm included on the agenda of high tech organizations as NASA or ESA. The application developer may meet different design tasks to be evolved. As the case, the design to be evolved could be: a program, a model of hardware or the hardware itself. Algorithms that run outside the reconfigurable hardware, mainly feature the actual EHW state of the art, but also some chip level attempts were done. It is important to understand that evolutionary circuit design and evolvable hardware (EHW) are two different and distinct approaches. Evolutionary circuit design performs the evolution (the design) of a single circuit. The aim is typically to design novel implementations that are better (in terms of area, speed, power consumption) than conventional deigns and/or to design circuits with additional features such as fault-tolerance, testability, polymorphic behaviour, that are difficult to design by conventional methods. Evolvable hardware (EHW) involves an EC responsible for continual adaptation. EHW is applied to highperformance and adaptive systems in which the problem specification is unknown beforehand and can vary in time. Editorial

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 3,

N째 2

We are indebted to the authors and reviewers for their efforts, outstanding contributions and assistance in the preparation of this special issue. We would like to express our sincere gratitude to Prof. Janusz Kacprzyk from the Polish Academy of Sciences the President of IFSA for encouraging and giving us the opportunity to edit this special issue. In the first paper, Oltean, Hintea, Sipos are focussed on new a method for analogue circuit design optimisation. Design objectives can be expressed in a flexible manner by using fuzzy sets. Neuro-fuzzy systems (universal approximators) are used to model the complex multi-variable and non-linear circuit performances. The exploration of the large, multidimensional solution space in quest for the optimal solution is done by an efficient and robust genetic algorithm avoiding local minima. This was tested and validated by application to the design optimisation of a CMOS amplifier. The second paper by Negoita, Sekanina, and Stoica is an overview on the Evolvable Hardware (EHW) the exciting and rapidly expanding industrial application area of the Evolutionary Computing (EC), of the Genetic Algorithms especially. An overview is made on the outstanding technological support making possible the implementation of system adaptation in hardware. Different kind of programmable circuits arrays are introduced. The most known EC based methods for the EHW implementation are described. A main part of this paper deals with some concrete elements of the EHW design, including the current limits in evolutionary design of digital circuits. Practical concluding remarks for the practitioners with regard to future perspectives of the area are an integrant part of this paper. Csipkes G, Hintea, Csipkes D, Rus, Festila and Fernandez-Canque deal with a reconfigurable and programmable analog low pass filter for low voltage wireless applications. The proposed filter may be used for channel or band selection in multi-mode receiver front ends employing direct frequency conversion. The circuit has been synthesized using state variables and leapfrog OTA-C techniques, features programmable order, digitally variable frequency parameters and wide linear range. The fundamental OTA cell has been implemented with fully balanced second-generation current conveyors, suitable for operating with low supply voltages. The functionality of the design has been demonstrated by transistor level simulations based on an 180 nm digital CMOS technology. A linear Support Vector machine classifier is proposed in this paper. In SVM linear classifiers architectures based on multiplying laws the main building blocks are multipliers. Festila, Szolga, Groza, Hintea, Cirlugea treat this approach. Using a model consisting of a compound of two inverse non-linear functions develops the multiplying and weighting cells. The procedure is fitting the VLSI implementation by use of simple nonlinearized standard log-domain or DA cells that compensate each other nonlinearities to obtain an extended domain of operation. Current-mode ELIN (externally linear internally nonlinear) design is used for its low voltage, low power and high speed characteristics. The resulted parallel-serial classifier was simulated taking into account real parameters of transistors in BICMOS technology. In the fifth paper, Kirei, Topa, Neag, Onet are focused on some aspects regarding the design process of the low-IF receivers. They present an I/Q imbalance image interference - compensation algorithm based on neural networks that is suitable for low-IF receivers. The standard solution using a complex LMS adaptive filter, which separates the desired, and image signals is limited in that the recovered signal remains affected by the I/Q imbalance, a drawback that is corrected by the proposed method. The functionality, convergence and stability of the neural network based filter are demonstrated through extensive computer simulations.

Editors: Mircea Gh. Negoita KES International, 2nd Floor, 145-157 St John Street, London, EC1V 4PY, United Kingdom; E-mail: m.negoita@hotmail.com Sorin Hintea Faculty of Electronics, Telecommunications and Information Technology, The Technical University of Cluj-Napoca, 26-28 GeorgeBaritiu Street, 400027 Cluj-Napoca, Romania; E-mail: Sorin.Hintea@bel.utcluj.ro

Editorial

2009

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 2

2009

ANALOG CIRCUIT DESIGN BASED ON COMPUTATIONAL INTELLIGENCE TECHNIQUES Gabriel Oltean, Sorin Hintea, and Emilia Sipos

Abstract: This paper presents a new method for analog circuit design optimization. Our approach turns to good account the advantages offered by computational intelligence techniques. Design objectives can be expressed in a flexible manner using fuzzy sets. This way appears the possibility to consider different degrees for requirement achievements and acceptability degree for a particular solution. Neurofuzzy systems (universal approximators) are used to model the complex multi-variable and nonlinear circuit performances. These models satisfy two main requirements: high accuracy and low computation complexity. An efficient and robust genetic algorithm avoids local minima in its exploration of the large, multidimensional solution space in quest for the optimal solution. Keywords: analog circuit design, optimization, genetic algorithm, fuzzy sets, neuro-fuzzy systems.

1. Introduction The purpose of analog circuit design is to produce a sized circuit schematic starting from a set of circuit requirements. Given a circuit schematic and the circuit’s performance specifications, the sizes and biasing of all devices have to be determined such that the circuit meets the specifications at some optimal cost. This is a difficult and critical step for several reasons: 1) most analog circuits require a custom optimized design; 2) the design problem is typically underconstrained with many degrees of freedom; and 3) it is common that many (often conflicting) performance requirements must to be taken into account, and tradeoffs must be made that satisfy the designer [1]. Optimizations tools appear, naturally, as the key factor for the tremendous effort of finding the design parameters, which satisfy a complex, high-dimensional, multi-objective and multi-constrained problem [2]. An optimization algorithm for analog circuit design has three key components: formulation of the optimization problem, performance evaluation engine, and optimization engine. Research efforts on circuit synthesis involving a broad spectrum of computational intelligence (CI) techniques have begun to appear in the literature over the past few years. Fuzzy sets are used to formulate the objective functions, getting this way the possibility to consider different degrees for requirement achievements and acceptability degrees for a particular solution. One approach [3] - [6], is to consider that the membership degree μ represents the degree of fulfillment of the fuzzy

objective. A value μ=1 means that the objective is fully satisfied, while a value μ=0 means that the objective is not satisfied at all. This method has a disadvantage in the case of equality requirement, no information being available regarding the relation between the requirement and the actual performance. An accurate estimation of the circuit performances requires the use of complex models leading to an excessively large computation time in the iterative optimization process. One way to reduce the computation time is to use more simple models of circuit performances. In order to satisfy both main requirements (accuracy and speed), many researches proposed several CI-based methods to evaluate circuit performances. For example least-square support vector machines are involved in [7] - [9]. Fuzzy systems are very useful to model the circuit performances because they imply just a few simple mathematical operations and can model any complex, multivariable and nonlinear function at any level of accuracy. Such fuzzy models are used in [10] - [13]. The optimization engine (the way to compute new parameter values) is the heart of the optimization algorithm. It should be chosen so that the optimization will converge to an optimal solution in a reduced number of iterations. This task is not an easy one due to complex relations between design parameters and circuit performances. A parameter affects more than one circuit performance at a time, so when a parameter is modified to improve a performance it can worsen another. To search the whole solution space, a powerful global optimization technique should be considered. Genetic algorithms (GA) are based on the Darwinian principle of natural selection and the concepts of natural genetics. GAs have many desirable characteristics and offer significant advantages over traditional methods. They are inherently robust and have been shown to efficiently search large solution spaces containing discrete or discontinuous parameters and non-linear constraints, without being trapped in local minima. GAs do not require initial guesses or derivative information and have often found non-intuitive solutions to engineering problems. Genetic algorithms have already been employed in many CAD applications [10], [14] - [17]. The objective of this work is to develop an efficient algorithm based on computational intelligence techniques for design optimization of analog circuits. Our approach tries to combine the best qualities of these techniques: flexibility in formulation the objective functions and a known range of their values using fuzzy sets; accuracy and low computational complexity of circuit performance models based on neuro-fuzzy systems; and a powerful global optimization engine based on Articles

Journal of Automation, Mobile Robotics & Intelligent Systems

genetic algorithm. The reminder of the paper is organized as follows. We begin in Section 2 with an overview of the CI-based optimization algorithm used in the design optimization of analog modules. The techniques used in the key phases of the algorithm are presented here. Section 3 focuses on the utilization of our proposed algorithm and results obtained for design optimization of a CMOS amplifier. In the end, in Section 4, some conclusions and further research directions are discussed.

2. CI-based Optimization Algorithm Design optimization of an electronic circuit is a technique used to find the design parameter values (length and width of MOS transistors, bias current, capacitor values etc.) in such a way that the final circuit perfor-mances (dc gain, gain-bandwidth, slew rate, phase margin etc.) meet as close as possible the design requirements. As stated in [18] there are two basic modalities to deal with the analog design: knowledge based approaches and optimization based approaches. In the present paper we are centered on the last one. 2.1. Overview of the Optimization Algorithm The optimization algorithm begins with the formulation of optimization objectives and optimization problem, followed by the initialization of the design parameters. During iterations an evaluation engine computes the actual circuit performances based on the actual design parameter values. If the objectives are fulfilled, the solution consists in the set (or sets - in the case of a real multiobjective optimization) of the actual design parameter values and the algorithm is stopped. If not, new design parameter values are to be computed by the optimization engine and the optimization loop is covered once again. The novelty introduced in this paper is the utilization of different CI techniques in all phases of optimization algorithm, as it is shown in Fig. 1.

Fig. 1. CI-based optimization. 64

Articles

VOLUME 3,

N° 2

2009

Fuzzy sets are used to define the objective function in formulation of the optimization problem. Neurofuzzy systems address the performance evaluation issue (evaluation engine). Finally, in the optimization engine, a genetic algorithm is responsible for the evolution of the population to finally produce the (near) optimum solution. 2.2. Formulation of the Optimization Problem To solve a multiobjective optimization problem, as is the design optimization of analog circuits, the optimization problem can be formulated in one of the following two ways [14]: 1) As a single-objective, constrained optimization problem, where different performance objectives are combined to form a single scalar objective, and which produces one solution. 2) As a multiobjective optimization problem, where the concept of Pareto-optimality is used to produce multiple tradeoff solutions on a design decision surface. Usually, for an analog circuit optimization problem with three or more objectives, the first approach appears to be computationally cheaper than the second approach. As a consequence this paper takes the single-objective approach. To formulate the design objectives for a real design is not always a simple task. Often, it is not clear what precise values should be given to each objective. In fact, design objectives are often better expressed in real world terms than in precise numbers. The designers usually can accept a certain degree of fulfillment of the design objectives. The fuzzy techniques used to define the optimization objectives suppose the fuzzification of the requirements, getting this way the possibility to consider different degrees for requirement achievements and acceptability degrees for a particular solution. The authors proposed the utilization of fuzzy sets to define the objective functions in previous papers [19] and [20]. By contrast with the existing approaches where membership degree represents the degree of fulfillment, in our approach the membership degree μ represents the error degree in the fulfillment of the objective. A value μ=1 means the objective is not satisfied at all, while a value μ=0 means that the objective is fully satisfied. With this approach we know the range of possible values for objective functions as being [0, 1]. When the value of a certain objective function (unfulfillment degree - UD) is 0, we know that the corresponding requirement is fulfilled, no further effort being necessary to improve the associated performance. The membership degree μ represents the error degree in the fulfillment of the fuzzy objective. A value μ=1 means the objective is not satisfied at all, while a value μ=0 means the objective is fully satisfied. As an example, the requirements “greater or equal” and “equal” have the corresponding fuzzy objective functions presented in Fig 2. where: x - the vector of the design parameters; th fk - the k performance function;

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

- the k requirements; x* - the current value of the design parameters vector. The fuzzy objective functions are. (1) is the range of possible values for fk(x). indicates the error degree in accomplishing th the k requirement, so we will call it “unfulfillment degree” (UD). A value =0 means full achievement of fuzzy objective, while a value =1 means that the fuzzy objective is not achieved at all. This occurs when fk(x) takes an unacceptable value. Fig 2 shows the corresponding value of the unfulfillment degree for the current value of the variables vector . With this approach we know the range of possible values for objective functions as being [0, 1]. When the value of a certain objective function (UD) is 0, we know that the corresponding requirement is fulfilled, no further effort being necessary to improve the associated performance, as usually happens in typical minimization-type optimization. where

Fig. 2. Fuzzy objective functions: a)

N° 2

2009

iterative process (see Fig. 1) that requires a large number of performance evaluations. Analog circuits are difficult and time-consuming for a proper evaluation. Even in the case of basic characteristics of a simple circuit (amplifier gain, gain-bandwidth, slew rate etc) the performance in question can be a complex function of many parameters. In a realistic case, a performance model will be in general a non-linear function over a high dimensional space of circuit parameters [9]. Even a small cell requires a mix of ac, dc and transient analyses. An accurate estimation of the circuit performances requires the use of complex models leading to an excessively large computation time. One way to reduce the computation time is to use more simple models of circuits and devices.

; b)

The formulation of the multiobjective optimization problem became: Find x that minimizes (2) where n is the number of requirements. For our single-objective optimization approach, we have to combine the individual objective functions into a cost function by means of a weighted sum: Fig. 3. Modeling procedure for circuit performance. Find x that minimizes

(3)

where wk is the relative preference or weight associated th with the k objective function. The requirement of the optimization, to satisfy all the objectives at the end of the optimization run, suggests that all the different objectives be weighted equally. On the other hand, for a given problem, some of the objectives may be more difficult to attain than some others. Thus, if a classification among the objectives is possible on grounds of relative difficulty of attainment, one would like to give higher numerical weights to the difficult objectives than the others. 2.3. Evaluation Engine The design process of an electronic circuit is an

Fuzzy systems are very useful to model the circuit performances because they imply just a few simple mathematical operations and can model any complex, multivariable and nonlinear function at any level of accuracy. The author synthesized a method to build neuro-fuzzy models and used it for some analog modules [19], [21], [22]. These models are automatically built up using an input-output data set, using the ANFIS (Adaptive Neuro-Fuzzy Inference System) framework [23] to develop first order Takagi-Sugeno fuzzy systems. A common way to apply a learning algorithm to a fuzzy system is to represent it in a special artificial neuronal network (ANN) like architecture. In the ANFIS framework, six-layer architecture for ANN is used. ANFIS makes use of a mixture of back propagation to learn the premise paraArticles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

meters and least mean square estimation to determine the consequent parameters [23], [24]. The full modeling procedure is explained in Fig. 3. The ranges of the parameter values are established so that irrespective of the parameter values combinations, the circuit will operate in the desired region. For example, in an amplifier the transistors should be maintained in their active regions. The parameter set (the combination of the parameter values) should be chosen to be representative for the function to be modeled (covers the space of the parameters and embeds all the specific characteristics of the function). For each input vector (one combination of the parameter values), the function value has to be found, in our case by SPICE simulation. Two data sets, a training set and a checking set are generated. The training set is then subdued to a subtractive clustering procedure resulting an initial first order TakagiSugeno fuzzy system. Next, the initial fuzzy system is trained using ANFIS and the training and checking data sets. The resulting neuro-fuzzy model is tested from the accuracy point of view. If the accuracy is unacceptable, the procedure must be resumed by generating a new initial fuzzy system or even by determining new data sets. If the accuracy is acceptable, the modeling procedure stops and provides the desired fuzzy model of that circuit performance. The main advantage of fuzzy models is that there are no restrictions in the kind of functions that can be modeled, as far as neuro-fuzzy systems are universal approximators. 2.4. Optimization Engine The heart of the whole algorithm is the optimization engine. A genetic algorithm (GA) is responsible for the exploration of solution space in quest of the optimal solution. Generally, the best individuals of any population tend to reproduce and survive, thus improving successive generations [25]. However inferior individuals can, by chance, survive and reproduce. In our case, the individuals consist of different versions (same topology, but different parameter values), which can evolve until a solution is reached (in terms of requirements accomplishment). Central to all genetic algorithms is the concept of the chromosome. The chromosome contains all information necessary to describe an individual. In nature, chromosomes are composed of DNA. In a computer, a long binary or character string is used. Chromosomes are composed of genes for the various characteristics to be optimized and can be any length depending on the number of parameters to be optimized. Basically, in a genetic algorithm each chromosome is potentially a solution of the optimization problem. Encoding defines the way genes are stored in the chromosome and translated to actual problem parameters. A generic chromosome employed in our algorithm is shown in Fig. 4, where each gene represents a design parameter. The “alphabet” used in the representation of genes can, in theory, be any finite alphabet. Thus, rather than 66

Articles

N° 2

2009

use the binary alphabet of 1 and 0, one can use an alphabet containing more characters and numbers [26].

Fig. 4. Format of a generic chromosome. Because the design parameters are real variables we chose real numbers to compose our alphabet. This way our population individuals are real valued vectors, rather than bit strings, thus simplifying the development of GA implementation. The underlying procedure of our GA is shown in Fig. 5.

1. Population initialization 2. Fitness assignment 3. Selection 4. Recombination 5. Mutation 6. Reinsertion 7. Loop to step 2 until optimal solution is found

Fig. 5. GA procedure. Seeding the population with random values, with a uniform probability, commonly does population initialization. It is sometimes feasible to seed the population with “promising” values that are known to be in the hyperspace region relatively close to the optimum [26]. Our implementation uses random at uniform initialization. Each individual in the selection pool receives a reproduction probability depending on the own objective value and the objective value of all other individuals in the selection pool. This fitness is used for the actual selection step afterwards. Our implementation uses rankbased fitness assignment with linear ranking [27]. Consider Nind the number of individuals in the population, Pos the position of an individual in this population (least fit individual has Pos=1, the fittest individual Pos=Nind) and SP the selective pressure. For example, in the case of a minimization-type optimization problem first position is allocated to the individual with highest value of the objective function. The fitness value for an individual is calculated as: (4) Linear ranking allows values of selective pressure in [1.0, 2.0]. For the selection our approach uses the roulettewheel method. This is a stochastic algorithm and involves the following technique. For each individual a selection probability is computed as:

(5)

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

The individuals are mapped to contiguous segments of a line, such that each individual's segment is equal in size to its selection probability. A uniformly distributed random number is generated and the individual whose segment spans the random number is selected. The process is repeated until the desired number of individuals is obtained (called mating population). This technique is analogous to a roulette wheel with each slice proportional in size to the fitness. Recombination produces new individuals in combining the information contained in two or more parents (parents - mating population). This is done by combining the variable values of the parents. Depending on the representation of the variables different methods must be used. For our real valued variables the intermediate recombination method was chosen. Offspring are produced according to the rule [27]:

N° 2

2009

(7)

The range of mutation is given by the value of the parameter r and the range of the variables. The parameter k (mutation precision) defines indirectly the minimal possible step-size and the distribution of mutation steps inside the mutation range. The smallest relative -k 0 mutation step-size is 2 , the largest 2 =1. By changing these parameters (r and k) very different search strategies can be defined. Our GA uses the pure reinsertion scheme: produce as many offspring as parents and replace all parents by the offspring. Every individual lives one generation only.

3. Design Optimization of a CMOS Amplifier (6)

where

represent the variable of the offspring, represent the variable of the first parent, while represent the variable of the second parent. The scaling factor is chosen uniformly at random over an interval for each variable anew. Intermediate recombination is capable of producing any point within a hypercube slightly larger than that defined by the parents. By mutation, individuals are randomly altered. Mutation of real variables means that randomly created values are added to the variables with a low probability. Thus, the probability of mutating a variable (mutation rate) and the size of the changes for each mutated variable (mutation step) must be defined. The probability of mutating a variable is inversely proportional to the number of variables (dimensions). The more dimensions one individual has, the smaller is the mutation probability. Different papers reported results for the optimal mutation rate. In [28] it is shown that a mutation rate of 1/n (n: number of variables of an individual) produced good results for a wide variety of test functions. That means that per mutation only one variable per individual is changed/mutated. Thus, the mutation rate is independent of the size of the population. The size of the mutation step is usually difficult to choose. The optimal step-size depends on the problem considered and may even vary during the optimization process. It is known that small steps (small mutation steps) are often successful, especially when the individual is already well adapted. However, larger changes (large mutation steps) can, when successful, produce good results much quicker. Thus, a good mutation operator should often produce small step-sizes with a high probability and large step-sizes with a low probability. Such an operator [27] was considered for our algorithm:

The proposed CI-based design optimization algorithm is implemented in the Matlab environment. It accepts three types of requirements: “greater than”, “equal”, and “smaller than”. The user should provide numerical values, types and weights for all the requirements. We used our algorithm for the design optimization of a CMOS simple operational transconductance amplifier (SOTA), shown in Fig. 6.

Fig. 6. Simple operational transconductance amplifier. The design parameters of the circuit are the dimensions of the transistors (W/L) and the bias current Ib. The input transistors Q1 and Q2 must be identical, therefore (W/L)1=(W/L)2 resulting the first parameter (W/L)12=(W/L)12. The transistors Q3 and Q4 (active load) must be paired, resulting (W/L)3=(W/L)4, so our second parameter will be (W/L)34=(W/L)34. For the current mirror, Q5 and Q6, we consider the current (Ib) equal trough both transistors so (W/L)5=(W/L)6. In order to keep a minimal area, we have taken W=L so we obtained our third parameter (W/L)56=(W/L)56. The fourth and final parameter is Ib. As performances, the important ones were considered: voltage gain Av0, unity gain bandwidth GBW, and common mode rejection ratio CMRR. Applying the previously described procedure, we built the neuro-fuzzy models of circuit performances with a set Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

of 850 data pairs (700 training pairs and 150 checking pairs). The design optimization is illustrated here for the set of requirements presented in Table 1, for equal weighted objective functions. The optimization was run several times, for a population of 100 individuals. The algorithm proved to be a robust one, always a solution being found that fulfills all the requirements. Different number of iterations is necessary to search for the optimum solution depending on the initial population and on the evolution process. Table 1 gives the final performances of the circuit after four different optimization runs, while Table 2 shows the solutions (the values of the design parameters). The solutions appear to be slightly different from each other. At a closer look we can see that the values of the design parameters are calculated with two decimals. In practical implementations these values should be rounded toward some discrete values, so we can consider that our resulted solutions are in fact small variations around one solution - our solutions are near optimum solutions. Table 1. Requirements and performances optimizing SOTA in four optimization runs.

NÂ° 2

2009

Fig. 8. Dynamic behavior of performances in run1. For the first optimization run (run1), the dynamic behavior of the algorithm is presented in Fig. 7. In the first iterations (up to 10), due to a high diversity of individuals, a new population does not contain always a fittest individual than in the previous population (minimum value of the cost function increases). On the other hand, the population as a whole is improved continuously, the average value on the entire population of the cost function decreasing in time. As the population improves during evolution, all individuals moves toward the optimal solution, decreasing both the minimum and average values of cost function. The evolution of all performance functions during optimization is presented in Fig. 8.

4. Conclusions Table 2. Solutions optimizing SOTA in four optimization runs.

A new computational intelligence-based optimization algorithm for analog circuit design was presented. The proposed algorithm was used to optimize the design of a CMOS simple transconductance amplifier, with very promising results. In a reduced number of iterations (63 to 94) it was able to always find an optimal solution, regardless the initial starting point (initial population), proving its robustness. The multiobjective optimization problem, specific to the analog circuit design, was reformulated as a single-objective optimization. For each optimization run the proposed algorithm produces one optimum solution. A further research direction is to use a real multiobjective optimization method to produce solutions on the Pareto front.

AUTHORS Gabriel Oltean, Sorin Hintea*, and Emilia Sipos Technical University of Cluj-Napoca, C. Daicoviciu Street, 15, Cluj-Napoca, Romania. E-mails: {Gabriel.Oltean, Sorin.Hintea, Emilia.Sipos}@bel.utcluj.ro. * Corresponding author

References [1]

Fig. 7. Minimum and average cost function evolution for run1 . 68

Articles

R.A. Rutenbar, G.G.E., Gielen, J. Roychowdhury, â&#x20AC;&#x153;Hierarchical Modeling, Optimization, and Synthesis for System-level Analog and RF Designsâ&#x20AC;?, In: Proc. of the IEEE, vol. 95, issue 3, 2007, pages 640-669.

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

[2]

[3]

[4]

[5]

[6]

[7]

[8]

[9]

[10]

[11]

[12]

[13]

[14]

[15]

M. Barros, J. Guilherme, N. Horta, “GA-SVM Optimization Kernel Applied to Analog IC Design Automation”. In: Proc. of 13th IEEE International Conference on Electronics, Circuits and Systems, ICECS06, 2006, pp. 486-489. B.Sahu, A.K. Dutta, “Automatic Synthesis of CMOS Operational Amplifiers: A Fuzzy Optimization Approach” . In: Proc. of 15th Intl. Conf. on VLSI Design, 2002, pp. 366-371. M. Fares, B. Kaminska, “FPAD: A Fuzzy Nonlinear Programming Approach to Analog Circuit Design”, IEEE Trans. on Computer-Aided Design of Integrated Circuits and Systems, vol. 14, no.7, 1995, pp. 785-793. B.R.S. Rodriguez, M.A. Styblinski, “Adaptive Hierarchical Multi-Objective Fuzzy Optimization for Circuit Design”. In: IEEE International Symposium on Circuits and Systems, ISCAS '93, 1993, pp. 1813-1816. L. Fu, J. Fan, L. Yuanchun, T. Yantao, D. Yisong, “Study and Application of a Constrained Multi-objective Optimization Algorithm”. In: IEEE International Vehicle Electronics Conference, IVEC 1999, pp. 305-307. X. Ren, T. Kazmierski, “Performance Modeling and Optimization of RF Circuits Using Support Vector Machines”. In: Proc. of 14th International Conference on Mixed Design of Integrated Circuits and Systems, MIXDES07, 2007, pp. 317-321. F. De Bernardis, M.I. Jordan, A. Sangiovanni Vincentelly, “Support Vector Machines for Analog Circuit Performance Representation”, DAC 2003, Anaheim, California, USA, 1-58113-688-9/03/0006, 2003, pp. 964-969. T. Kiely, G. Gielen, “Performance Modeling of Analog Integrated Circuits Using Least-Squares Support Vector Machines”. In: Proc. of Design, Automation and Test in Europe Conference and Exhibition, vol. 1, 2004, pp. 448-453. S. Balkir, G. Dundar, G. Alpaydin, “Evolution Based Synthesis of Analog Integrated Circuits and Systems”. In: Proc. of the IEEE NASA/DoD Conference on Evolution Hardware, EH'04, 2004 , 0-7695-2145-2/04, pp. 26-29. G. Alpaydin, S. Balkir, G. Dundar, “An Evolutionary Approach to Automatic Synthesis of High-Performance Analog Integrated Circuits”, IEEE Transaction on Evolutionary Computation, vol. 7, no.3, June 2003, pp. 240-252. A. Torralba, J. Chavez, L.G. Franquelo, “A Collection of Fuzzy Logic-Based Tools for the Automated Design, Modeling and Test of Analog Circuits”, IEEE Micro, vol. 16, no. 4, 1996, pp. 61-68. N.D. Venkata, B.L. Evans, “An Automated Framework for Multicriteria Optimization of Analog Filter Design”, IEEE Trans. on Circuits and System-II: Analog and Digital Signal Processing, vol. 46, no. 8, 1999, pp. 981-990. A. Somani, P.P. Chakrabarti, A. Patra, “An Evolutionary Algorithm-Based Approach to Automated Design of Analog and RF Circuits Using Adaptive Normalized Cost Functions”, IEEE Transactions on Evolutionary Computation, vol. 11, no. 3, 2007, pp. 336-353. T. Eeckelaert, R. Schoofs, G. Gielen, M. Steyaert, W. Sansen, “Hierarchical Bottom-up Analog Optimization Methodology Validated by a Delta-Sigma A/D Converter Design for the 802.11a/b/g standard”. In: Design

[16]

[17]

[18]

[19]

[20]

[21]

[22]

[23]

[24]

[25]

[26]

[27]

[28]

N° 2

2009

Automation Conf., 43rd ACM/IEEE, 2006, pp. 25-30. K. Takemura, T. Koide, H.J Mattausch, T. Tsuji, “Analogcircuit-component optimization with genetic algorithm”. In: The 47th Midwest Symposium on Circuits and th th Systems, vol. 1, 25 -28 July, 2004, pp. I - 489-92. M. Taherzadeh-Sani, R. Lotfi, H. Zare-Hoseini, O. Shoaei, “Design optimization of analog integrated circuits using simulation-based genetic algorithm”. In: International Symposium on Signals, Circuits and Systh th tems, SCS2003, 10 -11 July, vol. 1, 2003, pages 73-76. R.L. Gieger, P.E. Allen, N.R. Strader, VLSI Design Techniques for Analog and Digital Circuits, McGraw-Hill Publishing Company, 1990. G. Oltean, “FADO - A CAD Tool for Analog Modules”. In: Proc. of the International Conference on "Computer as a Tool", EUROCON2005, Belgrade, ISBN 1-4244-0050-3, IEEE catalog number: 05EX1255C, 2005, pp. 515-518. G. Oltean, “Multiobjective Fuzzy Optimization Method”, Scientific Bulletin of the Politechnica University of Timisoara, Trans. on Electronics and Communications, vol. 49, issue 63, no. 1, ISSN 1583-3380, 2004, pp. 220-225. G. Oltean, S. Hintea, Doris, Lupea, “A Fuzzy Optimization Engine for Analog Circuit Design”. In: Proc. of th the 8 International Conference MIXDES2001, Zakopane, Poland, 21st-23rd June 2001, pp. 77-82. G. Oltean, C. Miron, M. Crasi, M. Carlugea, “Evaluation of the Analog Circuits Performances Using Fuzzy Models”, Scientific Bulletin of the Politechnica University of Timisoara, Trans. on Electronics and Communications, vol. 47, issue 61, no. 1, ISSN 1583-3380, 2002, pp. 12-17. R.J.-S Jang, “ANFIS, Adaptive-Network-Based Fuzzy Inference System”, IEEE Transaction on System, Man, and Cybernetics, vol. 23, no. 3, 1993, pp. 665-685. C. Tran, A. Abraham, L. Jain, Decision Support Systems Using Intelligent Paradigms, eprint arXiv:cs/0405052, 2004. Reis, Cecilia, J.A.T. Machado, J.B. Cunha, E.J.S. Pires, “Evolutionary computation in the design of logic circuits”. In: IEEE International Conference on Systems, Man and Cybernetics, 7-10 Oct. 2007, pp. 1664-1669. R. Eberhart, Y. Shi, Computational Intelligence. Concepts to Implementations, Elsevier, Morgan Kaufman Publisher, ISBN 978-1-55860-759-0; 2007. H. Pohlheim, GEATbx Introduction to Evolutionary Algorithms: Overview, Methods and Operators, version 3.7 (November 2005), 2005, http://www.geatbx.com/. H. Mühlenbein, D. Schlierkamp-Voosen “Predictive models for the breeder genetic algorithm in continuous parameter optimization”, Evolutionary Computation, vol. 1, issue 1, ISSN:1063-6560, 1993, pp. 25-49.

Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

N째 2

2009

ADAPTIVE AND EVOLVABLE HARDWARE AND SYSTEMS: THE STATE OF THE ART AND THE PROSPECTUS FOR FUTURE DEVELOPMENT Mircea Gh. Negoita, Lukas Sekanina, Adrian Stoica

Abstract: This paper is an overview on the Evolvable Hardware (EHW) - the exciting and rapidly expanding industrial application area of the Evolutionary Computing (EC), of the Genetic Algorithms especially. The content of the work has the following structure: the first part includes generalities on industrial applications of EC, and the importance of EHW in this frame; the second part presents the outstanding technological support making possible the implementation of system adaptation in hardware. Different kind of programmable circuits arrays are introduced. The third part tackles the most known EC based methods for EHW implementation; the fourth part deals with some concrete elements of the EHW design, including the current limits in evolutionary design of digital circuits. The last part is focused on some concluding remarks with regard to future perspectives of the area. A list of references used in this work was inserted at the end. Keywords: Evolvable Hardware (EHW), Evolutionary Design, Reconfigurable Hardware, Field Programmable Analogue Arrays (FPAA).

1. Introduction Nevertheless, it is hardware implementation of the most benefit for the society and indeed most revolutionizing application of EC by leading to the so-called EHW. These new EC based methodologies make possible the hardware implementation of both genetic encoding and artificial evolution, having a new brand of machines as a result. This type of machines is evolved to attain a desired behavior that means they have a behavioral computational intelligence. There is no more difference between adaptation and design concerning these machines, these two concepts representing no longer opposite concepts. A dream of technology far years ago currently became reality: adaptation transfer from software to hardware is possible by the end. Much more, the electronics engineering as a profession was radically changed: the most based on soldering assembling manufacturing technologies are largely replaced now by programming circuitry-based technologies, including EHW technologies. 1.1.

Definitions, General Consideration and Classification of EHW A definition of the EHW may be as follows: a subdomain of artificial evolution represented by a design methodology (consortium of methods) involving the application of EC to the synthesis of digital and analogue electronic circuits and systems. A more agreed defini70

Articles

tion among the practitioners might be: EHW means programmable hardware that can be evolved [2]. But some members of the scientific community acting in the area consider the term evolutionary circuit design more descriptive for EHW features. Much more, another term is used nowadays for the same work - evolvare concerning to this evolvable ware with hardware implementation. This leads to a future perspective of using the term bioware concerning to a possible evolving ware with biologic environments implementation. Even some other environments are seen as possible evolvable media: wetware real chemical compounds are to be used as building blocks or nanotechnology - relied on molecular scale engineering. This new design methodology for the electronic circuits and systems is not a fash ion. It is suitable to the special uncertain, imprecise or incomplete defined realworld problems, claiming a continuous adaptation and evolution too. An increased efficiency of the methodology may be obtained by its application in the soft-computing framework that means in aggregation with other intelligent technologies [3; 4] such as fuzzy logic (FS) and neural networks (NN), Artificial Immune Systems (AIS), evolutionary algorithms (EA). The reason of EHW using the above mentioned type of application is relied on the main advantage over the traditional engineering techniques for the electronic circuit design, namely the fact that the designer's job is very much simplified following an algorithm [5] with a step sequence as below: STEP 1 - problem specification - requirements specification of the circuit to be designed specification of basic (structural) elements of the circuit; STEP 2 - genome codification - an adequate (genotypic) encoding of basic elements to properly achieve the circuit description; STEP 3 - fitness calculation - specification of testing scheme used to calculate the genome fitness; STEP 4 - evolution (automatically generation of the required circuit) - generation of the desired circuit. Despite of the above-mentioned advantage of EHW methodology, its disadvantage is not to be neglected: (sometimes) suboptimal results are got over those ones of the classical methods that remain preferable in case of well-defined problems. The designer himself is involved by acting directly during the first three steps, while the fourth step is an automatically generation of the circuit. The flow modality of both step 3 and step 4 leads to same categorizing classes criteria for EHW (see [5]). The practical EHW classification [6], can be adopted by: hardware type; controller type; Objective Function.

Journal of Automation, Mobile Robotics & Intelligent Systems

Based on the type of hardware and type of changes, EHW is classified as follows: - Reconfigurable digital hardware (RDH) - means digital hardware that changes quasi-instantly (short transition time compared with to processing time) and directly (single transition, no meaningful inter-mediary transition states) from the initial to the final configuration; - Morphable digital hardware (MDH) - means digital hardware that changes from one complete configuration to another through a set of intermediate states, each having a functional role; - Reconfigurable analog hardware (RAH) - is analog hardware with a dynamics of changes that is similar to RDH from the initial to final configuration; - Morphable analog hardware (MAH) - is analog hardware that changes gradually without switches; - Adjustable/Tunable/Parametric HW (ATPHW) - is hardware in which the changes influence the parameters of a function.

Classification by controller type leads to: Evolutionary hardware - EHW in which the controller employs evolution - any algorithms; Embryonic hardware - EHW in which the controller employs a embryonic - initiated growth mechanism.

The classification by Objective Function is a more general one, covering all kind of Adaptive Systems, not the EHW only: -

External adaptation - featured by adaptive behavior in the presence of stimuli originating in the surrounding environment; Internal adaptation - featured by adaptive behavior in the presence of a disturbance located in the system itself; Darwinian adaptation - featured by adaptive behavior when the response is directed toward modifying the object; Singerian adaptation - adaptive behavior when the response is directed towards modifying the environment; Smart System (SS) - a system aware of its state, operation and changing environment. It can predict what will happen at a certain future. This knowledge can lead to adaptation; Smart adaptive system (SAS) - can adapt to a changing environment, a similar setting without being â&#x20AC;&#x153;portedâ&#x20AC;? to it, adapt to a new/unknown application.

Dramatic changes happen in the relation between hardware and the application environment, and this in case of malicious faults or need for emergent new functions that claim for in-situ synthesis of a totally new hardware configuration. EHW is suitable for flexibility and survivability of autonomous systems as that ones developed by NASA JPL. EHW survivability means to maintain functionality coping with changes in hardware characteristics under the circumstances of adverse environmental conditions as for example: temperature variations, radiation impacts, aging and malfunctions. EHW

VOLUME 3,

NÂ° 2

2009

flexibility means the availability to create new functionality required by changes in requirements or environment. The application developer may meet different design tasks to be evolved. As the case, the design to be evolved could be: a program, a model of hardware or the hardware itself. Algorithms that run outside the reconfigurable hardware, mainly feature the actual EHW state of the art, but also some chip level attempts were done. The path from chromosome to behavior data-file is different in case of intrinsic and extrinsic (see more in [1]). EHW Evolution in Simulation has some typical features and parameters value as follows [5]: computationally intensive (640,000 individuals for about 1000 generations); runs over tens of hours, expected about 3 min in 2010 on desktop PC for experiments with netlists of about 50 nodes; SPICE scales badly, namely time increases nonlinearly with as a function of nodes in netlist, in about a subquadratic to quadratic way; no existing hardware resources allow porting the technique to evolution directly in hardware (and not sure will work in hardware). See in [1] more details on the huge advantage of the on- chip - versus CAD/synthesis tools. 1.2. The Technological Support of EHW The appearance of programmable integrated circuits, especially their new generation - field programmable gate arrays (FPGAs) and most recently reconfigurable analogue arrays (FPAAs) and field-programmable interconnection circuits (FPICs) or configurable digital chips at the functional block level, (open-architecture FPGAs) make possible for most companies to evolve circuits as the case would be. The appearance of reconfigurable analogue arrays (FPAAs) was crucial for EHW technological support. The analog reconfigurable hardware allows prevention or removal of essential fabrication mismatches and other refined technological problems by evolving circuits. The programmability of FPAAs is limited to just only allow configuration around op-amp level. But the applications require also for many interesting circuit topologies to be evolved below the op-amp level. This application requirement led to another kind of programmable (reconfigurable) hardware relied on evolutionoriented devices that have some advantages over FPAA, namely: can reprogram many times, can understand what's inside and are featured by a flexible programmability. This is the so called custom made EHW-oriented reconfigurable hardware. See in [1] more details regarding the types of custom made EHW-oriented reconfigurable hardware.

2. The Structure of an EHW System EC Based of EHW Implementation The language for programming reconfigurable hardware must define: an alphabet, expressing choices of cells and, a vocabulary/grammar, expressing the rules of in terconnect. An EHW system, any is its destination, either for demonstrations, prototype experiments or real time implementation, must be structured in from of two main components: the reconfigurable hardware (RH) and the reconfigurable mechanism (RM) [7]. Regarding the practical ways of implementing an evolvable system, Articles

Journal of Automation, Mobile Robotics & Intelligent Systems

real-world applications requirements are toward a reliable solution featured by compactness, low-power consumption and autonomy. The evolution on JPL SABLES (Stand-Alone Board-Level Evolvable System) [8] is a solution that proved to be effective in various applications, see Fig. 1.

Fig. 1. The simplified bloc diagram of JPL SABLES [14].

Fig. 2. The information flow between DSP and FPTA in JPL SABLE. The main integrated components by the JPL SABLES are: the transistor-level RH - a JPL FPTA, and the RM - the evolutionary algorithm, implemented by a TI DSP acting as a controller for reconfiguration. The information flow and implementation of JPL SABLES is featured - see Fig. 2 - both by autonomy and speed in providing on-chip circuit reconfiguration: about 1000 circuit evaluations are performed per second. Another parameters of JPL SABLES performance are the followings: 1-2 orders of magnitude reduction in memory and about 4 orders of magnitude improvement in speed compared to systems evolving in simulations. The final aim of EC techniques development and of their silicon implementation is to create architectures towards an artificial brain, a computer having its own ability of reasoning or decision making, but also being able of emergent functionality creation, or having the possibility of self-creation and evaluation of its own structure. The main types of EHW architectures with intrinsic EC logic elements are: embryological architectures; emergent functionality architectures; evolvable fault tolerant systems; parallel evolvable architectures (of Higuchi type) [1]. Other practical considerations regarding the concept of EHW architectures and hardware GA implementation, including the engineering compromise between the performances and the GA architectural complexity is to be found in [9]. Gate Level Evolution in FPGAs at the intrinsic level was applied in [10]. In the Functional Level Evolution the circuits are composed of high-level functional blocks such as adders and multiplier [11]. Incremental Evolution was applied in order to evolve circuits of high complexity. A divide-and conquers approach for the evolution of systems (also known under the name of “increased complexity evolution”) is introduced in [12]. EHW Architectures at Functional Level. This aim is feasible on a specialized dedicated FPGA architecture 72

Articles

VOLUME 3,

N° 2

2009

proposed by Higuchi [13] and called F PGA. This architecture is relied on a network of switch settings that allow the basic cells inside the device to be connected as the case. A basic cell in this kind of FPGA is called Programmable Floating Unit (PFU) because of its availability of performing a large function variety: adding, subtracting, multiplication, cosine and sine using floating point num2 bers. A F PGA programming word is a variable length chromosome containing both the PFU's function programming and the crossbar witch settings.

3. Evolutionary Design of Digital Circuits: Where Are Current Limits? It is important to understand that evolutionary circuit design and evolvable hardware (EHW) are two different and distinct approaches [14]. Evolutionary circuit design performs the evolution (the design) of a single circuit. The aim is typically to design novel implementations that are better (in terms of area, speed, power consumption) than conventional deigns and/or to design circuits with additional features such as fault-tolerance, testability, polymorphic behavior, that are difficult to design by conventional methods. Evolvable hardware (EHW) involves an EC responsible for continual adaptation. EHW is applied to high-performance and adaptive systems in which the problem specification is unknown beforehand and can vary in time [15], [16]. The main issues in the evolutionary circuit design are: the bias in the design method; the chromosome size versus complexity of circuits; the fitness calculation as a bottleneck; the level of innovation [14]. It is important to explore the relation between the size of chromosome and the complexity of evolved circuits. Long chromosomes imply large search spaces that are difficult to search. Computing resources that are currently available determine the size of the search space that can be explored. This is known as the problem of scalability of representation. In case of evolving the combinational circuits, the evaluation time of a candidate circuit grows exponentially with the increasing number of inputs. Hence, the evaluation time becomes a main bottleneck of the evolutionary approach even if the circuit consists of a few components. This is the problem of scalability of evaluation. In order to reduce this time of evaluation, only a subset of all possible input vectors can be utilized. It is impossible to measure the level of innovation in an evolved circuit; the level of innovation does not depend on the approach utilized to evolve the circuit and on the complexity of the evolved circuits. The Gate-Level Evolution can produce circuits of complexity up to about 100 gates. The Application-specific Encoding (computer & swapping medians or sorting networks in FPGA) can be used to obtain more complex circuits. The Functional Level Approaches produce circuits of an unlimited complexity and depends on the complexity of components used as building blocks. Development Approaches produce arbitrarily complex solutions using chromosomes of a short length. The advantage is that the area on a chip is minimized. The limitation is that the obtained solutions are featured by regularity; this causes redundancy. Transistor Level evolution produces only small digital circuits. These elementary circuits are

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

implemented using a large range of unconventional approaches. The complexity of evolved circuit only partially depends on the size of the chromosomes. It mainly depends on the size of objects encoded in the chromosome or handled by instructions encoded in the chromosome (see the developmental approaches), Fig.3, after [14].

NÂ° 2

2009

popsize - the population size; gnrs - the average number of generations to evolve the circuit. Some interpretations of different evolutionary design methods are as follows [14]: - the evaluation of a 21-input median circuit is the most time consuming; - pxg = popsize x gnrs, is a maximal parameter for CGP although the complexity of these gate-level circuits is not so high; surprisingly, pxg decreases with the complexity of evolved circuits; - the â&#x20AC;&#x153;time of evolutionâ&#x20AC;? = pxg x test-vect; this parameter is very similar for almost all circuits. This is important, because the time of evolution indicates how much time/resources the designers are able to invest to the evolutionary design process.

4. Concluding Remarks

Fig. 3. The complexity of the evolved circuits versus the size of chromosomes.

The most spectacular EC application in the CI framework is the (EHW). It opened a revolutionary eve in technology and in social life development by its radical implications on engineering design and automation. A dream of humanity became reality - the systems adaptivity was implemented (transferred) by EHW from software to hardware. A drastic time saving way from design to real world application of intelligent hardware is used, no more difference exists between design and adaptation concerning EHW based machines having a behavioral computational intelligence. Electronics engineering was fundamentally changed as a profession by using EHW custom design technologies instead of soldering based manufacturing. A survey of the state of the art of evolutionary design of digital circuits was made. The investigation was mainly focused on level of complexity of the evolved circuits with respect to the size of the search space. The practical conclusion is that innovative results can be obtained by all mentioned approaches. There is no reason to prefer some approaches.

5. Annex

Fig. 4. Evolved circuits complexity versus the configuration bitstream. Another relevant view is made by the comparison of computational effort considering circuits that have similar size of the configuration bitstream [14], see Fig. 4, where the chromosomes are of a size in the range of 512 -868. The legend of this table is as follows: in - the number of inputs; out - the number of outputs; chromsize the size of chromosome; gates - the number of gates the circuit is composed; testvect - the number of test vectors used in the fitness function to evaluate the circuit;

Field Programmable Analogue Circuits for Evolvable and Adaptive Hardware Field Programmable Analogue Arrays( FPAA). The appearance of reconfigurable analogue arrays (FPAAs) was crucial for the technological support required by companies involved in electronics research and development as well as in manufacturing. The analog reconfigurable hardware allows prevention or removal of essential fabrication mismatches and other refined technological problems by evolving circuits as the case. Analog reconfigurable hardware has actually a huge weight in EHW environment, see Fig. A.1, where the actual hardware platforms for EHW development and their authors are enumerated (Negoita and Stoica, 2004). FPAA reconfigurability relies on switched capacitors, typical to the main providers as Pilkington, Motorola see - Motorola MPAA020 - or Anadigm (see Fig. A.2). Zetex TRAC provides a Totally Reconfigurable Analog Circuit structured in form of 20 cells, each being an opamp with a small reconfigurable network. Such a cell performs any of the following analog operations: add, negate, subtract, multiply, pass, log, antilog, rectify, or baArticles

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 3,

N° 2

2009

sic inverting op-amp for use with external components.

Fig A.1. The main hardware platforms for EHW development and their authors/providers. Lattice ispPAC10 - see Fig. A.3 - contains four programmable analog modules and a programmable system. This circuit is an EHW analog reconfigurable hardware nd th that can be configured to implement 2 and 4 order active LP and BP filters in the 10 kHz-100 kHz range. High complex architectures are featuring some analog reconfigurable hardware of the Anadigm FPAA family, as for example AN220E04.

Fig. A.2 A typical FPAA switched capacitor configuration. Field Programmable Transistor Arrays (FPTA) produced by NASA JPL. The programmability of FPAAs is limited to only allow configuration around op-amp level. But the applications require also for many interesting circuit topologies to be evolved below the op-amp level. This application requirement led to another kind of programmable (reconfigurable) hardware relied on evolution-oriented devices that have some advantages over FPAA, namely: can reprogram many times, can understand what's inside and are featured by a flexible programmability. This is the so-called custom made EHW-oriented reconfigurable hardware. Some briefly comments are to be made regarding the types of custom made EHW-oriented reconfigurable hardware (Negoita and Stoica, 2004):

Articles

Fig.A.3 The block diagram of Lattice ispPAC10. JPL PTAs are chips of reconfigurable hardware at transistor level, both analogue and digital; FPTA chip of Heidelberg University is an array of 16x16 transistors, performing a programmability in connectivity and channel length; JPL'98 FPTA-0 chip is a programmable transistor array cell with 24 programmable switches, a sufficient number for meaningful circuit topologies. All three terminals of a transistor cell are connected via switches to expansion terminals. The chromosomes give the value HIGH-LOW of the switches (not only ON-OFF). JPL'2001 FPTA-2 chip is a second generation reconfigurable array chip, a programmable array of transistor array cells implementing an evolution-oriented reconfigurable architecture, featured by a NESW interconnection amongst 64 integrated cells (an 8x8 matrix of reconfigurable cells), each of the cells having 44 transistors. It is the first chip integrating reconfigurable processing circuitry with sensing: an array of 16x8 photodetectors distributed within the cells is also integrated on this chip. FPTA-2 chip is able of receiving 96 analog/digital inputs and provides 64 analog/digital outputs; it is the first FPMA (Field Programmable Mixed-signal Array). PAMA (Programmable Analog Multiplexer Array) chip is an analog platform based on analog multi-plexers/demultiplexers (Santini, Zebulum et al. 2001). The multiplexers/demultiplexers are fixed elements that perform the interconnections of the different discrete electronic devices that can be plugged into the board. The platform performs intrinsic evolution of analog circuits through a RM - represented by a GA. Each gene configures the select input signals of a particular analog multiplexer. A multifunction I/O board is connected to the PC bus to perform the A/D conversion and the chromosome download. The control bit

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

strings (GA chromosomes) are downloaded to the RH. The circuit evaluation runs as follows: circuit responses are compared against specifications of a desired response, this comparison being followed by a circuits ranking based on how close they come to satisfying the target. PAMA provides a practical environment to evolve generic analogue circuits based on discrete components, without the need of simulators. This is in fact a very useful prototype platform allowing a larg number of component terminals and evolution of a great number of circuits. The circuit evaluation speed is featured by 6 minutes to evolve a certain circuit. Other strong features is that PAMA platform provides protection against illegal configuration that may damage electronic components and confers the possibility to analyse circuits which have been evolved, due to access to individual circuit elements with test equipment. An innovative powerful Programmable System-on-Chip (PSoC) is the family PSoC™ CY8C25122/CY8C26233/ CY8C26443/CY8C26643 of configurable mixed-signal arrays provided by Cypress Connects. PSoC is a chip, reconfigurable device that replaces the components of a multiple MCU-based system components. A chip of this kind includes configurable analog and digital peripheral blocks, a fast CPU, Flash program memory and SRAM data memory in a range of convenient pin-outs and memory sizes.

[5]

[6]

[7]

[8]

[9]

[10]

ACKNOWLEDGMENTS The support of the EHW Group at NASA JPL was crucial for tackling research objectives surveyed in this work that was suggested by the KES International Advisory Board.

AUTHORS nd Mircea Gh. Negoita* - KES International, 2 Floor, 145157 St John Street, London, EC1V4PY, United Kingdom. E-mail: m.negoita@hotmail.com. Lukas Sekanina - Faculty of Information Technology, Brno University of Technology, Bozetechova 2, 612 66 Brno, Czech Republic . E-mail: sekanina@fit.vutbr.cz. Adrian Stoica - NASA Jet Propulsion Laboratory (JPL), USA, 4800 Oak Grove Drive MS 303-300, Pasadena, CA 91109. E-mail: adrian.stoica@jpl.nasa.gov. * Corresponding author

References [1]

[2]

[3]

[4]

Negoita Gh. M. (ed.), What is Evolvable Hardware, a Lecture, KES 2008 Post-Doc School on EHW and AHS, Zagreb, Croatia, September 2008. Torresen J., “Evolvable Hardware The Coming Hardware Design Method?”, In: Kasabov, (ed.), Neuro-fuzzy Tools and Techniques, Springer: Heidelberg, 1997. Negoita M.Gh., Neagu C.D., Palade V., Computational Intelligence. In: Engineering of Hybrid Systems, Springer; Heidelberg, 2004. Negoita M.Gh., “A Modern Solution for the Technology of Preventing and Alarm Systems: EC in EHW Implemen-

[11]

[12]

[13]

[14]

[15] [16]

N° 2

2009

tation”. In: Proceedings of The Second Dunav Preving International Conference on Preventing and Alarm Systems, Belgrade, 1997, pp. 201-209. Negoita M.Gh., Stoica A., “Evolvable Hardware (EHW) in the Framework of Computational Intelligence-Implications on Engineering Design and Intelligence of Autonomous Systems”. In: Tutorial Course, KBCS 2004, Hyderabad, India, December 2004, pp. 4-52. Stoica A., Radu A., “Adaptive and Evolvable Hardware A Multifaceted Analysis”. In: Arslan, T., Stoica, A., et al. (eds.) Proceedings of 2007 NASA/ESA Conference on th th Adaptive Hardware and Systems, Edinburgh, UK, 5 -8 August, 2007, pp. 486-496. Stoica A., Zebulum R.S., et al., “Evolving Circuits in Seconds: Experiments with a Stand-Alone Board-Level Evolvable System”. In: Proceedings of The 2002 NASA/ th th DoD Conference on EHW, Alexandria, Virginia, 15 -18 July, 2002, pp. 67-74. Zebulum R.S., Keymeulen D., Duong V., Guo X., Ferguson M.I., Stoica A., “Experimental Results in Evolutionary Fault-Recovery for Field Programmable Analog Devices”. In: Proceedings of The 2003 NASA/DoD Confeth th rence on Evolvable Hardware, Chicago, Illinois, 9 -11 July, 2003, pp. 182-188. Mihaila D., Fagarasan F., Negoita M.Gh., Architectural Implications of Genetic Algorithms Complexity in Evolvable Hardware Implementation. In: Proceedings of the European Congress EUFIT 1996, September, Aachen, Germany, vol. 1, 1996, pp. 400-404. Thompson A., Layzell P., Zebulum S., “Exploration in Design Space: Unconventional Electronics Design Through Artificial Evolution”, IEEE Transactions on Evolutionary Computation, vol. 3, vol. 3, 1999, pp. 167196. Murakawa M., et al., “Evolvable Hardware at Functional Level”. In: Ebeling, W., Rechenberg, I., Voigt, H.-M., Schwefel, H.-P. (eds.) PPSN 1996. LNCS, vol. 1141, 1996, Springer: Heidelberg (1996) pp. 62-71. Torresen J., “A Scalable Approach to Evolvable Hardware”, Genetic Programming and Evolvasble Machines, vol. 3, vol. 3, 2002, pp. 259-282. Higuchi T., “Evolvable Hardware with Genetic Learning”. In: Proceedings of IEEE International Symposium th on Circuits and Systems, ISCAS 1996, Atlanta, USA, 13 May 1996. Sekanina L., “Evolutionary Design of Digital Circuits: Where are Current Limits?” In: Stoica A., Arslan T. (eds.) Proceedings of First NASA/ESA Conference on Adaptive th th Hardware and Systems, Istanbul, Turkey, 15 -18 June, 2006, , pp. 171-177. Sekanina L., Evolvable Components: From Theory to Hardware Implementation, Springer: Heidelberg, 2004. Torresen J., Bake W.J., Sekanina L., “Recognizing Speed Limit Sign Numbers by EHW”. In: Raidl G.R., Cagnoni S., Branke J., Corne D.W., Drechsler R., Jin Y., Johnson C.G., Machado P., Marchiori E., Rothlauf F., Smith G.D., Squillero G. (eds.) EvoWorkshops 2004, LNCS, vol. 3005, 2004, Springer, Heidelberg, pp. 682-691.

Articles

Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 3,

NÂ° 2

2009

A HIGHLY LINEAR LOW PASS FILTER FOR LOW VOLTAGE RECONFIGURABLE WIRELESS APPLICATIONS Gabor Csipkes, Sorin Hintea, Doris Csipkes, Cristian Rus, Lelia Festila, Hernando FernĂĄndez-Canque

Abstract: This paper presents a reconfigurable and programmable analogue low pass filter for low voltage wireless applications. The proposed filter may be used for channel or band selection in multi-mode receiver front ends employing direct frequency conversion. The circuit has been synthesized using state variables and leapfrog OTA-C techniques, features programmable order, digitally variable frequency parameters and wide linear range. The fundamental OTA cell has been implemented with fully balanced secondgeneration current conveyors, suitable for operating with low supply voltages. Transistor level simulations based on a 180 nm digital CMOS technology have demonstrated the functionality of the design. Keywords: software radio, reconfigurable filters, state variable synthesis, current conveyors, dynamic bias currents.

1. Introduction The widespread of mobile communication systems have led to an unprecedented diversity of functions that the radio access interfaces must integrate. Each of these functions may be associated with a particular communication standard, defined by a precise set of specifications, including channel bandwidth and spacing, carrier frequency, data modulation, analog-digital signal processing and hardware requirements. The ultimate goal of the industry is the development of the software defined radio (SDR) technology that gives the context for the implementation of different analogue and digital reconfigurable building blocks. In the first acceptance the software defined radio has been envisioned as a signal processing chain in which the data converters are placed directly in the vicinity of the antenna, allowing all the specific radio functions to be realized in the digital domain. In spite of the advantages offered by the ideal architecture, the technology induced limitations concerning the consumption of a data converter capable of sampling the signal at the carrier frequency with sufficiently high dynamic range, may render a real implementation impossible. Therefore, the ideal model must be extended in order to include some analogue interface circuitry [1]. One of the most widely used radio interfaces, considered mainly due to its simplicity and well-understood behaviour, is the so-called zero-IF or direct conversion architecture shown in Fig. 1. Its simplicity makes the direct conversion architecture a good candidate for the implementation of reconfigurable radio interfaces [2].

Articles

Fig. 1. Typical reconfigurable direct conversion receiver architecture. In a direct conversion receiver the radio frequency signal is converted directly to the base band by multiplication with a complex local oscillator signal having the same frequency with the RF carrier. The LO signal is generated in quadrature to suppress image components and to allow the correct demodulation of the wanted signal. The advantages of the direct conversion approach over the classical heterodyne architecture are straightforward: there is no need for multiple frequency translations, amplifying and filtering stages, while the channel selection is simply done by means of low pass filters. A survey of the existing literature shows that the reconfigurable circuits, specifically channel or band select filters within the analogue interface, are key building blocks of a practically feasible SDR transceiver. Although filter design is a very well known area of VLSI circuit development, only few theoretical studies have been published on low voltage reconfigurable filters [3],[4]. The remainder of this paper describes a modular and easily scalable approach to reconfigurable and programmable analogue filter design. The proposed low pass filter is intended for SDR built upon zero-IF analogue interfaces and operates from a single 3 V supply voltage, compatible with modern digital CMOS fabrication technologies.

2. Reconfigurable state variable OTA-C filters State variable filters stand out among analogue filter architectures due to their low sensitivity performance, inherited from their doubly terminated passive LC prototypes. The implementation based on operational transconductance amplifiers (OTA-s), connected in open loop configurations, allows the operation at high frequencies [9]. Furthermore, the modular structure creates the premise for easy topological reconfiguration, a prerequisite for the design of fully reconfigurable filters. The versatility in operation is achieved by creating a fundamental module that can be simply cascaded in order to implement higher order filters. The modular structure is based on the circuit template that implements the signal flow graph associated with a generalized LC ladder prototype. The graph corresponding to

Journal of Automation, Mobile Robotics & Intelligent Systems

a generalized passive LC ladder is shown in Fig. 2. The block diagram has been obtained by writing Kirchoff's theorems and Ohm's law for the nodes and branches of the LC ladder. The dummy resistance R, usually sized to be equal to the termination resistances of the ladder, transforms branch currents into voltages allowing voltage only state variables [5].

VOLUME 3,

N째 2

2009

scaling. If the characteristic resistance of the ladder is R=RS=RL=1/Gm and ft is the desired corner frequency, the capacitances in the active implementation may be calculated according to the equations (1) [5]. From the sizing equations it can be seen that the corner frequency may be adjusted without modifying the shape of the frequency response by changing Gm. Furthermore, the adjustment of the corner frequency is independent on the filter order.

(1)

Fig. 2. The signal flow graph corresponding to a generalized LC ladder filter. Filters designed for communications systems exhibit mainly a low pass or band pass frequency response. In particular, when the imposed operating frequencies are not prohibitive and quadrature signal paths are available, band pass filters may also suppress the image signal in the intermediate frequency stage. In these cases the band pass response is obtained from the low pass transfer function by performing a linear frequency transformation. Therefore, the impedances in the generalized ladder should be particularized in order to accommodate first of all with a low pass response. In this case the odd order transversal components are capacitances and the even order longitudinal components are inductances as shown on the passive low pass prototype in Fig. 3.

where C2k+1(n) and L2k(n) are the normalized transversal capacitances and longitudinal inductances of the passive low pass prototype. The complete reconfiguration of the low pass filter implies freely changing the frequency parameters, the filter order and the approximation while dynamically adjusting the current consumption. A careful examination of the filter topology for two consecutive orders may be done when considering the corresponding termination networks for odd and even order implementations shown in Fig. 5. The purpose of the comparison is the identification of a reconfigurable module, simply cascaded in order to obtain a generalized, variable order filter.

Fig. 3. An odd order doubly terminated low pass ladder filter. The resulting implementation of the low pass OTA-C filter for an arbitrary odd order n is shown in Fig. 4 [5]. The even order filter can be derived through a similar procedure. Fig. 5. A comparison between odd and even order termination networks.

Fig. 4. Schematic of a fully balanced odd order OTA-C filter. The values of the inductances and capacitances can be calculated after denormalization and impedance

The comparison shows that the fundamental module can be implemented around two transconductor cells with switches that connect or disconnect the given module depending on its position in the cascade. Additionally, the negative feedback created around the second OTA in every cell effectively implements an active resistance. The last module in the cascade will have the feedback path activated corresponding the load resistance RL of the ladder [6]. The resulting reconfigurable module is illustrated in Fig. 6.

Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

N째 2

2009

3. The Transconductance Amplifier

Fig. 6. The reconfigurable filter module. A simple cascade connection of identical modules, as many as required by the highest desired filter order, controlled by a decode logic, leads to a fully reconfigurable low pass filter implementation. The typical interconnection of several modules in a cascade configuration is shown in Fig. 7.

The most important high frequency analogue filter implementation techniques encountered in the literature are based on OTA-C structures. However, transconductance amplifiers are known to exhibit a limited range for the transconductance parameter, low-to-average linear range and relatively high current consumption. Classical linearization methods make use of the non-linear transistor equations and various circuit topologies in order to effectively cancel odd order non-linearities, while even order harmonics may be reduced by fully balanced designs [9]. The circuit described in this work is an alternative to a wide linear range, fully balanced OTA, built around second-generation current conveyors. Current conveyors play the same role in current mode signal processing as opamps in voltage mode circuits, namely they hold virtual ground for the wanted signal. Their versatility in applications is mainly defined by the configuration of the terminals. A classical method to

Fig. 7. A cascade of several identical modules to form a filter with variable order. When the filter order is decreased by one, the longith tudinal switches Slong in the n module of the cascade are turned OFF separating the unit from the rest of the ladder. Meanwhile, the switches Sfb are also turned OFF and the disconnected OTA cells are forced in a power-down state. The new outputs of the filter will be Op(n-1) and Om(n-1) of the previous module. Additionally, the switches Sfb of the termination module n-1 must be turned ON in order to shift the load resistance to the output of the lower order filter. The position of the switches for a filter with consecutive orders n-2, n-1 and n is given in Tab. 1.

build a transconductance amplifier with a CCII cell is to use the Y terminal as a voltage input. The input voltage is then copied to the X terminal. The voltage VX determines a current through a resistor connected between the X terminal and the ground. This current is then copied to the Z terminal, used as current output [7]. Mirroring the circuit and connecting a resistor between the X terminals leads to the OTA block diagram shown in Fig. 8.

Table 1. Switch states for three consecutive filter orders n-1, n-2 and n. Order

Switch Module n-2

Module n-1

Module n

n-2

Slong Sfb

ON ON

OFF OFF

n-1

Slong Sfb

ON OFF

ON ON

OFF OFF

Slong Sfb

ON OFF

ON ON

Articles

Fig. 8. Block diagram of the programmable transconductance OTA. The overall transconductance of the circuit will be equal to the reciprocal of the passive resistance R. The programmable output stage is essentially a weighted gain current amplifier permitting the adjustment of the transconductance and of the filter corner frequency. The fully balanced implementation requires a common mode control circuit (CMFB) that sets the DC voltages at the high impedance output terminals [10].

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

In the ideal case the equivalent transconductance Gm is 1/R. However, real current conveyors exhibit non-zero parasitic resistances of the X terminal, which must be added twice to the passive resistance R when calculating the transconductance [7]. Furthermore, RX is often frequency dependent and highly non-linear. The solution that reduces the relatively high parasitic resistance is to use negative feedback. There are various low input resistance conveyor implementations proposed in the literature. In order minimize the number of stacked transistors between supply rails and allow low voltage operation, the majority of these circuits are using a differential amplifier on the negative feedback path [7]. The filter design presented in this paper is built around Liu's conveyor [8]. Liu's CCII is essentially an unbuffered opamp with Miller compensation, whose second gain stage is used as a current mirror for copying the X terminal current to the Z output. Its main advantages are the simple structure, low current consumption and the potential to operate at supply voltages as low as 1.2 V. The latter option makes the designs compatible with modern, deep submicron CMOS technologies. A special care must be taken of the opamp stability through an adequate compensation. The lead-lag type of compensation, implemented with the resistor RC in series with the capacitor CC, yields good results in insuring the stability and in extending the range of the conveyor operating frequencies [7]. The schematic of Liu's CCII is shown in Fig. 9.

N째 2

2009

The calculations yield

(2)

The calculations show that the negative feedback helps reducing the small signal RX to hundreds or even tens of W-s. However, the non-linear expression of RX becomes increasingly important when considering the large signal behaviour. It can be demonstrated that the incremental current flowing through the X terminal exhibits a parabolic dependence on the voltage at the X terminal, according to (3), where a is the gain of the differential amplifier, bp is the intrinsic transconductance of the p-channel transistor M1 and IB is the bias current of the X terminal input stage [7]. (3) This equation shows that the non-linear RX introduces even order harmonics and it may be considered the main source of nonlinearity in CCII based circuits. Using fully balanced circuit architectures effectively cancel even order non-linearities. Therefore, when the differential signal paths are matched, the OTA has the potential for very linear operation along with relatively low current consumption. The Z terminal output resistance may also be enhanced by using cascoded transistors. The structure of the fully balanced OTA is presented in Fig. 11.

Fig. 11. The fully balanced OTA with cascode output stage.

4. Simulation Results

Fig. 9. Simplified schematic of Liu's current conveyor. The X terminal resistance may be determined from the small signal model of the input stage, illustrated in Fig. 10.

The simulations performed on the filter at transistor level have the main goal to demonstrate the functionality of the circuit in the presence of the inherent non-idealities, compared to the concept level implementation in [6]. The main emphasis here lays on the performance indicators concerning frequency and order variability, non-linearity and a dynamically adjustable current consumption. The corner frequency programming strategy implies changing the transconductances through a binary weighted current amplifier in the CCII output stage according to (4). In this equation 4 MHz is the lowest achievable corner frequency, R is the termination resistance of the network and bi are the programming bits of the weighted current amplifier.

(4) Fig. 10. The small signal model used to calculate RX. Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 2

2009

Fig. 12. The filter magnitude response for different orders and corner frequencies. Fig. 12 shows the magnitude response of the filter, designed for the lowest possible corner frequency equal to 4 MHz, standard Butterworth approximation, orders 4, 5, 6 and 7, various programming codes and corner frequencies. The linearity of the filter has been extensively simulated for all the corner frequency-order combinations. The frequency of the input signal has been chosen 250 kHz, such that the lowest ten harmonics lay in the filter pass band. The worst-case total harmonic distortion was approximately – 53 dB for a 1 V peak-to-peak differential input sine wave. Fig. 13 shows the typical output spectrum of the filter and the higher order harmonics for the th 6 order configuration. The built in power down algorithm allows the dynamical reduction of the filter consumption when the order th is lowered. The complete 4 order filter draws 8.4 mA th from a single 3 V supply, while the 7 order configuration draws 13.6 mA. The consumption has been measured for the highest possible corner frequency, when all the branches of the programmable output stage are functional and correctly biased.

the circuit can be extended to implement a polyphase band pass response, adapting the transfer function to the needs of other receiver configurations, such as the low-IF architecture. The simulations performed at transistor level have proven the functionality of the reconfiguration concept. The filter achieves good performances in terms of linearity and features dynamically adjustable current consumption. In all the cases the consumption is lower compared to classical OTA-C implementations with similar linear range. ACKNOWLEDGMENTS The authors would like to thank Mentor Graphics for their support in the current research.

AUTHORS Gabor Csipkes*, Sorin Hintea, Doris Csipkes, Cristian Rus, Lelia Festila - Technical University of Cluj-Napoca, Str. Daicoviciu 17, 400027, Romania. E-mail: gabor.csipkes@bel.utcluj.ro Hernando Fernández-Canque - Glasgow Caledonian University, School of Engineering & Computing, Cowcaddens Road, Glasgow, G4 0BA. * Corresponding author

References [1] [2]

[3] th

Fig. 13. Distortion measurement of the 6 order filter – 1V peak-to-peak input at 250 kHz.

[4]

5. Conclusions The reconfigurable low pass filter described in this paper is suitable for integration into multi-mode receiver front-ends that employ a form of direct conversion architecture. The modular design is easily scalable, allowing the extension to higher orders and wider programming range of the corner frequency by simple replication of the fundamental reconfigurable module. Furthermore, 80

Articles

[5]

[6]

[7]

Mittola J., “The Software Radio Architecture”, IEEE Communications Magazine, 1995, pp. 26-38. Maurer L., et al., “On the Architectural Design of Frequency Agile Multi-standard Wireless Receivers”. In: 14th IST Mobile and Wireless Communications Summit, Dresden, June 2005. Selex Communications, Wideband Structural Antenna Operating in the HF Range, Particularly for Naval Installations; patent, 2006. “Reconfigurable SDR Equipment and supporting Networks. Reference Models and Architectures”, Wireless World Research Forum, white paper, 2004. Csipkes D., et al., “Synthesis method for state variable Gm-C filters with a reduced number of active components”, MIXDES 2003, pp. 292-297. Hintea S., et al., “On the Design of a Reconfigurable OTA-C Filter for Software Radio”. On: NASA/ESA Conference on Adaptive Hardware and Systems (AHS2007). K. Koli, K. Halonen, Cmos Current Amplifiers: Speed

Journal of Automation, Mobile Robotics & Intelligent Systems

[8]

[9] [10]

VOLUME 3,

N° 2

2009

Versus Nonlinearity, Kluwer, 2002 Liu S., Tsao H., Wu J., “CCII Based Continuous-Time Filters with Reduced Gain-Bandwidth Sensitivity”, IEE Proceedings-G, vol.138, 1991, pp. 210-216. Kardontchik J.E., Introduction to the Design of Transconductor-Capacitor Filters, Kluwer, 1992. Forghani-Zadeh H.P., Rincon-Mora G.A., “A Continuous, Low-Glitch, Low-Offset, Programmable Gain and Bandwidth Gm-C Filter”, 48th Midwest Symposium on Circuits and Systems, vol. 2, 7th-10th August 2005, pp. 16291632.

Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 2

2009

AN ANALOG LINEAR SVM IMAGE CLASSIFIER

Lelia Festila, Lorant Andras Szolga, Robert Groza, Sorin Hintea and Mihaela Cirlugea

Abstract:

A linear Support Vector machine classifier is proposed in this paper. In such SVM architectures based on multiplying laws the main building blocks are multipliers. We propose in this paper multiplying and weighting cells, developed by using a model consisting of a compound of two inverse non-linear functions. This procedure is suitable for VLSI implementation because it permits the use of simple nonlinearized standard log-domain or DA cells that compensate each other nonlinearities to obtain an extended domain of operation. Current-mode ELIN (externally linear internally nonlinear) design is used for its low voltage, low power and high speed characteristics. The resulted parallelserial classifier was simulated taking into account real parameters of transistors in BICMOS technology. Keywords: Support Vector Machine, analog multipliers, logdomain th domain, square-root domain, image classifier. b)

1. Introduction Support vector machines (SVMs) are a set of related supervised learning methods used for classification and regression [9]. A binary SVM classifier has to decide which of two classes C and C` an object belongs. It classifies positively the object if it belongs to the class C and negatively if it does not [1], [3]. The classifier is trained with positive and negative labeled vectors of features (objects or data points) belonging or not to the given class C. The training results consist of M support vectors (SVs) Xm, m=1,2…M, which include relevant features of the training vector set and also Lagrange - {-1,1} assigned to each coefficients am and labels ymC SV. Label ym is 1 if Xm belongs to class C and -1 if it belongs to C`. A vector X to be classified is given to the already trained classifier characterized by SVs Xm, m=1,…,M and their corresponding Lagrange coefficients and labels. (1) The classifier calculates a decision function. Its sign represents the label y to be assigned to the tested vector X, which is positive classified if y ³ 0. For example the linear classifier considered in this paper has the decision function based on a multiplying law and the label to be calculated is of the following form [4]: (2)

Articles

Fig. 1. Block diagram of a cellular parallel SVM classifier with multiplying law: (xk xkj), k = 1,…,N; j = 1,…,M; N - vector length; M - number of SV; a) with multiplying cells; b) with weighting cells. Examining relation (2) one can see that to implement such a SVM classifier we need multipliers as basic T functional unities. The X Xm operation is implemented by a matrix of multiplying cells as Fig. 1a shows. Each T product X Xm is then multiplied by its corresponding coefficient amym. If the resulted um, m=1,…,M signals are currents, they can be summed by simple connections. A constant current b is added then and finally a current comparator delivers the decision function y [5]. If a SVM classifier is dedicated to a specific application, SVs Xm and coefficients could be fixed parameters

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

set by design and the basic functional unities become simple weighting elements, as Fig. 1b shows. As SVM classifiers usually require a large amount of calculations, their VLSI implementations need high density, high speed and low power circuits.

NÂ° 2

2009

-1

One connects two F and F blocks as in Fig. 3. If the requirement V1=V2=V may be fulfilled, a current-mode weighting cell or a multiplier can be realized (Fig. 4):

if Fig. 4. Current-mode weighting cell or multiplier. Fig. 2. Proposed Log-Domain SVM classifier block diagram. Fig. 2 shows the block diagram of the serial-parallel classifier presented in this paper. The multiplier array has the structure of a column in the NxM array in Fig. 1. Support vectors are serially introduced and also coefficients T aiyi. To add continuously each vector product XXi to the partial sum a log-domain integrator is used. The circuit also contains at the output a signal comparator to classify the input vector of data. Because the main building blocks in the SVM architecture in Fig. 1 and Fig. 2 are multipliers or weighting elements we propose in this paper current-mode multiplying and weighting cells, developed by using a model consisting of a compound of two inverse non-linear functions. This procedure is suitable for VLSI implementation because it permits the use of simple nonlinearized cells that compensate each other nonlinearities to obtain an extended domain of operation - procedure characteristic for ELIN (externally linear-internally nonlinear) circuits [2]. We will also present the other component circuits for the SVM parallel classifier proposed in this paper and then the simulation results for an image recognition table.

If currents I1 and I2 can be set independently of voltages V1 and V2 , respectively, multipliers result. The type of the multiplier depends on the input port type. The input signal may admit one or two directions, respectively, but scaling currents I or voltages V are usually one-directional in the basic building blocks. Therefore, by using the block diagrams from Fig. 3, 1-Q or 2-Q multipliers result. For a four-quadrant multiplier an extra current source is needed, as Fig.5 shows as an example.

Fig. 5. Basic 4-Q current mode multiplier model: ix and ix are bi-directional input signals. Fig. 6a and 6b show some variants for implementing -1 block F with an inverting nonlinear block F. a)

2. Multiplier current-mode models -1 and schematics based on F-F functions if -1

2.1. Multiplying F-F models Let define an invertible function F: x ÂŽ y, where variables x and y are nondimensional, expressed by normalized voltages and currents respectively:

(3) Consider that the above functions can be implemented by two basic building blocks F, a nonlinear transcon-1 ductor, and F , a nonlinear transresistor represented in Fig. 1. a)

-1

Fig. 6. Implementation of function F with: a) a differential F block; b) an inverting F block. In the following we analyze examples of multipliers -1 based of the above-proposed F-F models in log-domain, or hyperbolic-domain. In square-root domain the model leads to an uncontrollable gain. 2.2. A Log Domain 4Q Multiplier Based on models given in Fig. 3-6, we developed in [5] log-domain multipliers with ln-exp building blocks. -1 In this case F and F functions are:

-1

Fig. 3. F (Fig. 1a) and F (Fig. 1b) building block symbols.

(4) Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

The basic circuits and their symbols are shown in Fig. 7. Input Ix in Ln block is onedirectional so the multiplier having the structure from Fig. 3 is a 1Q one. a)

N째 2

2009

Fig. 8. 4Q log-domain multiplier: a) block diagram; b) DC sweep simulation results.

The relation between in/out signals and the linearity of the characteristics proves the validity of models. 2.3. Multiplying cell realized with bipolar DAs a)

f) b)

Fig. 7. Log-Exp building blocks: a) ln circuit; b) symbol of ln block; c),d) exp circuits; e), f) symbol of exp circuits. The 4Q multiplier based on the model in Fig. 3 is presented in Fig. 8a. The input-output characteristics for i1C{[-9uA, 9uA] having i2 as parameter [-9uA,5uA,-1uA,1uA,5uA,9uA] are given in Fig. 8b. a)

Fig. 9. Bipolar differential amplifier used in the large signal -1 domain a) F block: th cell; b) F block: arcth cell. The large signal model of a DA in bipolar technology is described by a hyperbolic tangent function as Fig. 9a -1 shows. In Fig. 9b the connection implementing F function is given. The input signal in the F cell is . One can see that the requirement V1=V1=VT is fulfilled and a 2Q current mode multiplier may be realized using the model from Fig. 3. Fig. 10 shows the schematic of such a 2Q multiplier and iout=f(ix) characteristics for Iy=ct as a parameter. In Fig. 11a 4Q multiplier is shown considering the model from Fig. 4. Simulations have proved the validity of the models and also the linearity on the whole domain |ix|,|iy| < I. The same conclusions result if schematics from Figs 9-10 are realized with MOS transistors in weak inversion, -1 because functions F-F remain of the same type.

Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 2

2009

-1

Fig. 12. Square-root F and F cells a) F cell ; b) F cell. b)

-1

In the same figures the functions F and F are deduced on the base of the input output relations Fig. 11a and Fig. 12b. The circuit having the model in Fig. 3 is shown in Fig. 13. In this case considering Fig. 3, relation V1=V2 has to be fulfilled, that is: (4)

Fig. 10. 2Q multiplier a) schematics; b) in-out characteristics for Iy= ct and I=50μA.

One can see that the current gain Ai cannot be controlled. Therefore the current amplifier from Fig. 13 could only be used to realize weighted signals with fixed weights.

b) Fig. 11. 4Q multiplier for Iy= ct and Ix=100μA a) schematic; b) in-out characteristics. Weighting cells with MOSFETs saturated in strong inversion -1 Figures 12a and 12b present both F and F functions implemented by simple differential amplifiers (DAs) with saturated MOS transistors in strong inversion. a)

Fig. 13. Current amplifier for W2/L2 = 2*W1/L1 a) schematics; b) in-out characteristic.

3. SVM classifier components A SVM parallel-serial classifier having the architecture shown in Fig. 2 is proposed in BICMOS technology. The basic building blocks are: multiplying cells, integrator and comparator. Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

The variant proposed in this paper contains log-domain multipliers shown in Fig. 8 based on e and ln modules from Fig. 7. Because one has to sum sequential signals resulted by multiplying test vector X with each support vector Xm a log-domain integrator was used as Fig. 14 shows [6]. The final result y is given by a current comparator [11] in Fig. 15.

NÂ° 2

2009

The images to be tested are given in Fig. 17. The resulting output currents of the integrator corresponding to each tested image are shown in Table 2.

Fig. 17. Test images. Table 2.

Fig. 14. Log-Domain Integrator.

Fig. 15. Current comparator circuit.

4. Simulation Results Simulations have been performed taking into account applications in image classifications. For example a SVM was trained by software to identify the letter F in a 6x7 image. A number of 16 positive and 16 negative examples were considered in the training. The results are 16 SV's as Fig. 16 shows and also coefficients aiyi, i=1,2â&#x20AC;Ś6 given in Table1.

The input current for a black pixel was considered -1 mA and the value for the white one +1 mA; the bias current b is 4 mA. At the output of the inverting current comparator, the positive values of the decision function indicate the images close to the given reference image, in our case letter F. In Fig. 18 we present the signals from the outputs of the main building blocks, for a given test image, in this case, Tst1 from the Fig. 18. The following output signals are considered: the output current of the multiplier array, a), the output current from the second multiplier block, which performs the product between the first signal and the SV's coefficients presented in Table 1; b), the resulting current at the current summator log-domain block, c) and finally the output voltage of the current comparator d). The time necessary to realize the classification operation with the given number of SV's was 160 ms. a)

b) Fig. 16. SV's resulted after the training process. Table 1. Support Vector Coefficients.

Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

[3]

[4]

[5]

d) [6]

[7]

[8] [9] [10]

Fig. 18. Simulation results: a) output signal from the multiplier array; b) output signal of the second multiplier; c) output signal from the log-domain current summator; d) output signal from the current comparator. [11]

5. Conclusion Modularity, current programmability of the parallelserial SVM classifier proposed in this paper recommend it for standard cell design and real time operation in image classification tasks. ELIN design was extended in this paper to modular nonlinear circuits like multipliers. The general modular multiplier models use simple -1 nonlinear F-F cells that compensate each other nonlinearities. In this manner the large signal domain can be used extending the linear behavior of input-output characteristics. Replacing the F modules by simple nonlinearized bipolar or CMOS building blocks very simple multiplying/weighting cells may result. They are suitable to be used in VLSI neuron-like networks in particular SVM classifiers because of their simplicity, efficiency and good performance for a large domain of signal variations.

N° 2

2009

and Their Application to Companding Signal Processing”, IEEE Trans. on CS II, vol 44, no. 2, Feb. 1997, pp. 65-84. Gordan M., Doctoral Thesis: Applications of the Fuzy technics and the SVMs in the image processing (in Romanian), Technical University of Cluj-Napoca, 2004. Joachims T., Support Vector and Kernel Methods, Cornell University, Computer Science Department, SIGIR, 2003. Festila L., Groza R., Szolga L., Hintea S., “Log-Domain multipliers for VLSI architectures”, Scientific Bulletin of the “Politechnica” University of Timisoara, Timisoara, September 2006, pp. 121-125. Groza R., Festila L., Fazakas A., A Log-Domain Sum-mingAmplifier for Serial Signal Flows. Inter-Ing 2007, Tîrguth th Mures, IV 4-1IV4-4,15 -16 November 2007. Dominguez-Castro T., Rodriguez-Vazquez A., Huertas J.L., “High resolution CMOS current comparators”. In: Proc. 1992 European Solid State Circuits Conf., 1992, pp. 242-245. Joachims T., SVM Light Version: 6.01, University of Dortmund, Informatik, AI-Unit, 02.09.2004. http://en.wikipedia.org/wiki/Support_vector_machines Festila L., Szolga L., Cirlugea M., Groza R., “Analog Multiplying/Weighting VLSI Cells for SVM Classifiers”, International Conference on Knowledge-Based and Intelligent Information & Engineering Systems, ISSN 3029743, Zagreb, Croatia, 2008. Groza R., Festila L., Sorin H., Cirlugea M.., “Log-Domain binary SVM image classifier”, International Conference on Knowledge-Based and Intelligent Information & Engineering Systems, ISSN 3029743, Zagreb, Croatia, 2008.

AUTHORS Lelia Festila*, Lorant Andras Szolga, Robert Groza, Sorin Hintea and Mihaela Cirlugea - Basis of Electronics Department, Technical University of Cluj-Napoca. G.Baritiu 26-28 Street, Romania. E-mail: lelia.festila@bel.utcluj.ro. * Corresponding author

References [1]

[2]

Genov R., Chakrabartty S., and Cauwenberghs G., “Silicon support vector machine with on-line learning”, Int. J. Pattern Recognition Artificial Intell., vol. 17, no. 3, 2003, pp. 385-404. Tzividis Y., “Externally Linear, Time Invariant Systems Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

NÂ° 2

2009

I/Q IMBALANCE COMPENSATION ALGORITHM BASED ON NEURAL NETWORKS Botond Sandor Kirei, Marina Topa, Marius Neag, Raul Ciprian Onet

Abstract: This paper proposed an I/Q imbalance compensation algorithm based on neural networks, suitable for low-IF receivers. First, the low-IF receiver architecture and the phenomena of I/Q imbalance (also referred as image interference) are described. The standard solution - using a complex LMS adaptive filter, which separates the desired, and image signals - is limited in that the recovered signal remains affected by the I/Q imbalance; the filter proposed here corrects this drawback. The functionality, convergence and stability of the neural network based filter are demonstrated through extensive computer simulations. A sizing example is also given - deduction of the number of sample necessary in order to achieve a -60 dB image rejection along with the time domain behaviour of the resulting neural network. Keywords: low-IF receiver, I/Q imbalance compensation, image rejection, and neural networks.

1. Introduction In the last two decades, increased consumer interest in wireless communication devices has resulted in significant technical developments. Nowadays rigorous efforts are invested in development of multistandard receivers with low cost, single chip implementation, low power consumption, etc. The main contenders for title of most popular architecture for integrated receivers in this century have been so far the zero-IF and low-IF receivers. Both architectures are suitable for integration and have been used to develop complex SoC, combining analog front-ends and digital base-band signal processing on a single chip. However, each of these architectures has serious drawbacks, such as: for the low-IF receiver, nonidealities inherent to physical implementations result in amplitude and phase mismatches between the I and Q signal paths; thus the desired signal is degraded by interferences (â&#x20AC;&#x153;leakageâ&#x20AC;?) from the adjacent band signal, making it mandatory to use image rejection filters. The desired signal in zero-IF receiver is degraded by time variant (low-frequency) DC offsets caused by self-mixing and leakages between the local oscillator and the RF path; such errors are very difficult to eliminate or compensate for. Numerous image rejection algorithms and filters for low-IF receivers are described in the literature. An interesting solid state circuit solution has been recently proposed in [1]: an adaptive filter based on the sign detection LMS algorithm that allows for a simple hardware implementation at the cost of lower estimation accuracy. 88

Articles

Depending on the quantization noise, it can achieve an image to signal ratio (ISR) of approximately 60dB. Reference [2] presents a non-data-aided image rejection algorithm. Exploring the mutual independence property of the desired and image signals the measure of the interference can be determined and the initial signals restored. By using the simulation model presented in [2], one can see that the desired signal expression comprises a scaling factor that is phase mismatch dependent, thus making the filter vulnerable to phase errors. A similar problem limits the solution proposed in [3]: there a modified complex LMS filter is used in order to reject the image frequencies; the amplitude errors are corrected but the phase errors are not fully cancelled. Reference [4] shows how a small phase mismatch can reduce the ISR achieved by an LMS filter down to 30 dB. Other solutions are based on blind source (or signal) separation [5]. The earliest implementations make use of a calibration tone signal applied at the front-end of the low-IF receiver [6]. These solutions are becoming obsolete, as they required additional calibration time and hardware for the tone generation. The image rejection filter proposed here is based on the one described in [3]. As mentioned above, the major drawback of this implementation is its relatively large sensitivity to phase errors and mismatches between I and Q signal paths, highlighted by the presence of a phase mismatch dependent in the expression of the resulting wanted signal. The solution presented here aims at correcting this problem by using neural networks to implement the LMS adapting function, thus eliminating the scaling factor. Note that although there are several noise sources in a low-IF receivers (noise introduced by the analog mixer, quantization noise due to the analogdigital conversion, image interference, etc.), this paper analyses only the effect of image interference. In Section 2 and 3 the low-IF receiver architecture and the image interference - also referred as I/Q imbalance - are briefly presented. Section 4 describes the enhanced image rejection filter based on a neural network. For a better understanding, the combination of two neurons resulting in the filter studied in [3] and [4] is first presented. Then the newly proposed combination of four neurons, that realizes the enhanced image rejection, is described. Section 5 contains simulation results obtained in Simulink; last but not least, conclusions are drawn in Section 6.

2. The low-IF receiver architecture Fig. 1 presents the generic low-IF receiver architecture. The radio frequency (RF) signal captured by the

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

antenna is filtered by the band-select filter BPF and amplified by the low-noise amplifier LNA. Image cancellation can be achieved at this point, but requires narrow band filtering and thus increases significantly the complexity and cost of the device. The RF signal is down converted to an intermediary frequency (IF) by using a quadrature local-oscillator signal xLO; this way a complex low-IF signal is generated, which can be represented by in-phase and quadrature signals, usually named I and Q (I(t) and Q(t)). The IF signals are low-pass filtered by the LPFs then sampled by the ADCs, resulting in the I(n) and Q(n) strings. The base band demodulation and image rejection are done in the digital domain. It should be noted that some implementations - not discussed in this paper - use image-rejecting complex LPFs.

NÂ° 2

2009

conversion these errors will cause interferences between the I/Q paths; phase and amplitude mismatches between the I/Q signal paths (LPF and ADCs) are additional causes of such interferences. This effect is usually called I/Q imbalance. In order to simplify the mathematical expressions one can introduce the I/Q imbalance parameters, defined as follows: (6) The IF signal can now be expressed using the I/Q imbalance parameters: (7) where LPF stands for the low-pass filtering function. The IF signal spectrum is depicted in Fig. 2b. After the ADC sampling and conversion, the digitised IF signal is downconverted into the base band, yielding the following complex signals:

Fig. 1. The low-IF receiver architecture. (8) Let us consider that the RF signal r(t) at the input of the receiver is given by: (1)

Note that s(n) and i(n) are complex signals, as well.

where the carried signal z(t) is a combination of the desired signal s(t) and the interferer i(t) from the adjacent band: (2) The filtered intermediary frequency signal is sampled by the ADC resulting in: (3) The low-IF receiver will produce two output signals. The first one, d(n), is the intermediary frequency signal demodulated on the cosine carrier and filtered. The second signal, v(n), results from the intermediary frequency signal by demodulation with the sine carrier. Their expressions are:

(4)

Fig. 2. a) Spectrum of the RF signal. b) Spectrum of IF signal. c) Spectrum of the mixture signal d(n). d) Spectrum of the mixture signal v(n). e) Spectrum of desired signal s(n). f) Spectrum of interferer signal i(n). The signal d(n), depicted in Fig. 2c, contains the desired signal s(n) (spectrum shown in Fig. 2e and the conjugate of the interferer i*(n). Likewise v(n), depicted in Fig. 2d, contains the interferer signal i(n) (spectrum shown in Fig. 2f) and the conjugate of the desired signal s*(n). Therefore they are called â&#x20AC;&#x153;mixtureâ&#x20AC;? signals. The equations (6) can be written in compact matrix form as follows:

3. I/Q imbalance in low-IF receivers

(9)

In a real-life implementation the local oscillator signal, xLO, is affected by amplitude and phase errors than can be expressed by: (5) where g is the amplitude and j the phase errors. These errors have a slow variation in time, so during the signal processing they can be considered constant. At down-

4. Image rejection filter The signal flow chart in Fig. 3 corresponds to the operation of a standard LMS image rejection filter, such as the one proposed in [3]; it has two outputs, one for the desired and one for the interfering signal. In this case the neurons are not performing an LMS adaptation but an adaptive prediction operation. Thus one can write two Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

cost functions to force the outputs to represent the desired and interfered signals:

N째 2

2009

(15) while the complementary neurons aim to adapt in order to reach the following objective:

(10) (15) By substituting in equation (8) the desired signals r1(n) and r2(n) with the corresponding values of d(n) and v*(n) and the pair of input signal x1(n) and x2(n) with v*(n) and d(n) it results:

(11)

Fig. 4. Signal flow graph of proposed filter. Based on same considerations, the neuron coefficients for the recursive processes result:

Fig. 3. Signal flow graph of image rejection filter. For implementing a recursive process similar to LMS weight adaptation several considerations have to be made: 1. One pair of adapting neurons implies the usage of two weight values. 2. Because the neuron performs a prediction operation, the input signal x1(n) and x2(n) should be replaced by the errors e2(n), respectively e1(n), representing the values that are rejected. The resulting recursive process for the neurons coefficients is described by the following expressions:

(16)

Similarly, for the predicted signals one obtains the expressions: (17)

(12)

When w1 and w2 are adapted the following conditions are fulfilled:

From this point on the recovery of the desired signal consists in a gain correction given by the weight values; since it results:

(13) (18) For the predicted signals the corresponding expressions are: (14)

The imbalance parameters k1 and k2 still contain the phase error introduced in the mixing stage, but the signals are successfully separated. Equation (14) is equivalent to the result obtained in [3]. Fig. 4 presents the enhanced image rejection filters flow chart: one can observe that the filter contains four neurons, thus four cost functions will be processed. Two neurons are following the rules represented by equation 90

Articles

5. Simulations Since the presented image rejection filter is an adaptive one, stability, convergence and performance issues need to be carefully analysed. Extensive simulations have been run using Simulink models. In the example presented here the desired signal s(n) is an 8-QAM coded while the interferer (adjacent-channel) i(n) is 6-QAM. The d(n) and v(n) mixtures applied to the filter are obtained from the linear combination of s(n) and i(n) considering relative large values for the amplitude and phase errors: g=1.2 and j=10째 (see equations 6 and 9). Note that in

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

N째 2

2009

Fig. 5. a) Number of samples necessary to achieve a given ISR, as a function of the learning rate; b) ISR evolution in time for m=0.0005. nowadays-integrated radios the amplitude mismatch between the I/Q branch is 1-2% and the phase mismatch is well bellow 5째. Fig. 5a gives the number of samples necessary to reach a target ISR (-20, -40 and -60 dB) for different learning rates. From this graph an optimum value for the learning rate can be chosen, making a trade off between adaptation time and performance. By analysing these plots one can conclude that by decreasing the learning rate, the time to reach a given ISR increases exponentially. On the other hand, by decreasing the learning rate a higher precision filtering is achieved. The convergence of this system has not been proven mathematically but extensive simulations over a large range of conditions have shown that the filter always converges. Fig. 5b presents a typical example of such simulation: it shows the evolution of ISR from the first sample until the filter is adapted. The learning rate value

was set to m=0.0005 based on previous simulation results. One can observe that at the beginning the filter is not stabilized, it does not reject the image signals, and the ISR values are meaningless; after about 18 K samples the adaptation begins and the ISR value is decreasing monotonously; after approximately 70 K samples the filter hits a resting point, and the ISR tends to vary between two values. This is because the weight values of the neurons can move around the optimal solution with the freedom given by the learning rate.

6. Conclusions This paper proposed an I/Q imbalance compensation algorithm using neural networks, able to increase significantly the image rejection ratio of low-IF receivers. The new algorithm was based on an LMS filter solution proposed in the literature but solves a significant drawback of that filter, its sensibility to phase imbalances. This improvement was achieved by implementing an additional Articles

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

adaptive loop using neural networks. A Simulink model was developed and extensive simulations were run in order to demonstrate the effectiveness of the proposed algorithm and study its dynamic behaviour and convergence. They shown that the system converges over a wide range of conditions and it can provide ISRs better than 60 dB, even if the analog front-end introduces significant (larger than usual) amplitude and phase imbalances. There is a trade off between ISR and the learning rate/adaptation time. Further developments consist of implementing the proposed filter on an FPGA. The FPGA integration will be helped by the fact that the Simulink model was created so that it can serve like an RTL description of the filter. Another direction is to speed up the learning process by the usage of variable learning rate or other similar methods. ACKNOWLEDGMENTS This work was partly supported by the Romanian National University Research Council under Grant TD 428/2007 entitled Contributions to Designing and Implementing Integrated Polyphase Filters for Wireless Applications.

AUTHORS Botond Sandor Kirei*, Marina Topa, Marius Neag, Raul Ciprian Onet - Technical University from Cluj-Napoca, Department of Basis of Electronics, C. Daicoviciu. 15, 400020 Cluj-Napoca, Romania. E-mails: {botond.kirei, marina.topa, irina.dornean}@bel.utcluj.ro. * Corresponding author

References [1]

[2]

[3]

[4]

[5]

[6]

[7] 92

Lerstaveesin S., Song B.-S., “A Complex Image Rejection Circuit with Sign Detection Only”, IEEE Journal on Solid-State Circuits, vol. 41, no. 12, December 2006, pp. 2693 2702. Gil Gye-Tae, Kim Young-Doo, Lee Yong H., “Non-DataAided Approach to I/Q Mismatch Compansation in LowIF Receivers”, Transaction on Signal Processing, , vol. 55, no. 7, July 2007, pp. 3360-3365. Yu L., Snelgrove W. M., “A Novel Adaptive Mismatch Cancellation System for Quadrature IF Radio Receivers”, IEEE Trans. on. Circuits and Systems II: Analog and Digital Signal Processing, vol. 46, no. 6, June 1999, pp. 789-801. Kirei B. S., Dornean I., Topa M., “Image Rejection Filter Based on Complex LMS Filter for Low-IF Receiver”. In: Proceedings of IEEE - ELMAR 2008, vol. 1, Zadar, Croatia, th th 10 -12 September 2008, ISBN 978-953-7044-08-4, pp. 203-206. Windisch M., Fettweis G., „Blind I/Q Imbalance Parameter Estimation And Compensation In Low-If Receist vers”. In: IEEE Proceedings of 1 International Symposium on Control, Communications and Signal Processing (ISCCSP '04), Hammamet, Tunisia, March 2004. Glas J.P.F., „Digital I/Q Imbalance Compensations in a Low-IF Receiver”, IEEE Global Telecommunications Conference 1998 (GLOBECOM'98). The Bridge to Global Integration, vol. 3, 1998, pp. 1461 1466. Kiss P., Arias J., Li D., Boccuzzi V., “Stable high-order

Articles

[8]

[9]

N° 2

2009

delta-sigma DACs”. In: IEEE Proceeding of the 2003 International Symposium on Circuits and Systems, vol. 1, May 2003, pp. 985-988. Farhang-Boroujeny B., Adaptive Filters - Theory and Applications, John Wiley & Sons, 1998, ISBN 0-47198337-3, pp. 178-182. Haykin S., Neural Networks - A Comprehensive Foundation, Prentice Hall, 1999, ISBN 0-7803-3494-9, pp. 118-137.

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

NÂ° 2

2009

INFocus THE SPOTLIGHT on new n Robot Fish will Detect Sea Pollution British scientists from BMT Group Ltd developed carpshaped robot capable of detecting water pollution. Next year will start a trial of the first five robotic fish in the northern Spanish port of Gijon. It will be a part of a threeyear research project funded by the European Commission and co-ordinated by BMT Group. The first five fish are being built by Professor Huosheng Hu and his robotics team at the School of Computer Science and Electronic Engineering, University of Essex. The biomimetic robot mimics the movement of real carp and is equipped with tiny chemical sensors to sniff out potentially hazardous pollutants, such as leaks from vessels or underwater pipelines. It transmits the information back to the port's control centre using Wi-Fi technology. This is the first robotic fish able to navigate independently without any human interaction. It can return automatically to its hub Source of the image: www.bmt.org to be recharged when battery life (approximately eight hours) is low. Rory Doyle, senior research scientist at engineering consultancy company BMT Group, which developed the robot fish with researchers at Essex University, said there were good reasons for making a fish-shaped robot, rather than a conventional mini-submarine, hopes that the robotic fish will used in rivers, lakes and seas across the world. The robot fish will be 1.5 meters long - roughly the size of a seal, swim at a maximum speed of about one metre per second, and cost 20,000 pounds ($29,000) a piece. More information at http://www.bmt.org/News/?/3/0/510 n Automatic Beauty The new star of catwalks was born in Japan. Scientists from Japan's National Institute of Advanced Industrial Science and Technology presented the new version of HRP-4C during Tokyo Fashion Week in March. Standing at 186-cm tall and 43-kg with battery (so it has slimmed-down from an earlier 58 kg), cybernetic model has a shape designed to match the average Japanese woman; with humanlike skin and eyes, face and hair based on Japanese "anime" comics, is probably the most advanced humanoid robot in the world. Reports say that the robot, "kept looking surprised, opening its mouth and eyes in a stunned expression, when the demonstrator had asked it to smile or look angryâ&#x20AC;? - everything thanks to 30 motors spread throughout its body with an additional eight motors in its face. During the fashion show audience could observe - a constant look of surprise on HRP-4C's face. Second weak point is robot's bearing; however its movements are quite smooth, as well as constantly bent knees and unsteady gait make an impression of artificiality. As Reuters notes the robot's outfit is more like that of a storm trooper. Japanese model robot costs ca $200,000 (without a face and silver suit). The software will be available free - everybody will be allowed to program new behaviour's patterns. In our humble opinion HRP-4C should try walking in 12-inch platform heels, as flesh and bone models. Source: http://www.aist.go.jp/

Source of the image: AP In the spotlight

93 ?

VOLUME 3,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 2

2009

EVENTS SPRING-SUMMER 2009 May 07 – 08

ECT-2009 – International Conference on Electrical and Control Technologies 2009 – Kaunas, Lithuania. http://www.ktu.lt/ect

15 – 17

ICSPS'09 – International Conference on Signal Processing Systems, Singapore, Singapore. http://www.icsps.org

15 – 17

ICCDA'09 – International Conference on Computer Design and Applications, Singapore, Singapore. http://iccda.org

June 06 – 08

ICCEA'09 – International Conference on Computer Engineering and Applications, Manila, Philippines. http://www.iccea.org

24 – 26

ICIVC 2009 – International Conference on Image and Vision Computing, Paris, France. http://www.waset.org/wcset09/paris/icivc/

27 – 29

ICCNT'09 – International Conference on Computer and Network Technology, Chennai, India. http://www.iccnt.org

27 – 29

ICMEE'09 – International Conference on Mechanical and Electronics Engineering, Chennai, India. http://www.icmee.org

28 – 30

2 Engineering Summit – Montreux, Vaud, Switzerland http://www.engineeringgipfel.com/

July 14 – 16

August

The 7 International Conference on Field and Service Robots, Cambridge, Massachusetts, USA. http://www.rec.ri.cmu.edu/fsr09 nd

08 – 11

ICCSIT'09 – 2 International Conference on Computer Science and Information Technology 2009, Beijing, China. http://www.iccsit.org

26 – 28

ICBCSB'09 – International Conference on Bioinformatics, Computational and Systems Biology, Singapore, Singapore. http://www.waset.org/wcset09/singapore/icbcsb/

28 – 30

ICECS'09 – 2 International Conference on Environmental and Computer Science, Dubai, United Arab Emirates. http://www.icecs.org

28 – 30

ICCEE'09 – 2 International Conference on Computer and Electrical Engineering, Dubai, United Arab Emirates. http://www.iccee.org/

Events